15d4f98a2SYan Zheng /* 25d4f98a2SYan Zheng * Copyright (C) 2009 Oracle. All rights reserved. 35d4f98a2SYan Zheng * 45d4f98a2SYan Zheng * This program is free software; you can redistribute it and/or 55d4f98a2SYan Zheng * modify it under the terms of the GNU General Public 65d4f98a2SYan Zheng * License v2 as published by the Free Software Foundation. 75d4f98a2SYan Zheng * 85d4f98a2SYan Zheng * This program is distributed in the hope that it will be useful, 95d4f98a2SYan Zheng * but WITHOUT ANY WARRANTY; without even the implied warranty of 105d4f98a2SYan Zheng * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 115d4f98a2SYan Zheng * General Public License for more details. 125d4f98a2SYan Zheng * 135d4f98a2SYan Zheng * You should have received a copy of the GNU General Public 145d4f98a2SYan Zheng * License along with this program; if not, write to the 155d4f98a2SYan Zheng * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 165d4f98a2SYan Zheng * Boston, MA 021110-1307, USA. 175d4f98a2SYan Zheng */ 185d4f98a2SYan Zheng 195d4f98a2SYan Zheng #include <linux/sched.h> 205d4f98a2SYan Zheng #include <linux/pagemap.h> 215d4f98a2SYan Zheng #include <linux/writeback.h> 225d4f98a2SYan Zheng #include <linux/blkdev.h> 235d4f98a2SYan Zheng #include <linux/rbtree.h> 245a0e3ad6STejun Heo #include <linux/slab.h> 255d4f98a2SYan Zheng #include "ctree.h" 265d4f98a2SYan Zheng #include "disk-io.h" 275d4f98a2SYan Zheng #include "transaction.h" 285d4f98a2SYan Zheng #include "volumes.h" 295d4f98a2SYan Zheng #include "locking.h" 305d4f98a2SYan Zheng #include "btrfs_inode.h" 315d4f98a2SYan Zheng #include "async-thread.h" 325d4f98a2SYan Zheng 335d4f98a2SYan Zheng /* 345d4f98a2SYan Zheng * backref_node, mapping_node and tree_block start with this 355d4f98a2SYan Zheng */ 365d4f98a2SYan Zheng struct tree_entry { 375d4f98a2SYan Zheng struct rb_node rb_node; 385d4f98a2SYan Zheng u64 bytenr; 395d4f98a2SYan Zheng }; 405d4f98a2SYan Zheng 415d4f98a2SYan Zheng /* 425d4f98a2SYan Zheng * present a tree block in the backref cache 435d4f98a2SYan Zheng */ 445d4f98a2SYan Zheng struct backref_node { 455d4f98a2SYan Zheng struct rb_node rb_node; 465d4f98a2SYan Zheng u64 bytenr; 475d4f98a2SYan Zheng /* objectid tree block owner */ 485d4f98a2SYan Zheng u64 owner; 495d4f98a2SYan Zheng /* list of upper level blocks reference this block */ 505d4f98a2SYan Zheng struct list_head upper; 515d4f98a2SYan Zheng /* list of child blocks in the cache */ 525d4f98a2SYan Zheng struct list_head lower; 535d4f98a2SYan Zheng /* NULL if this node is not tree root */ 545d4f98a2SYan Zheng struct btrfs_root *root; 555d4f98a2SYan Zheng /* extent buffer got by COW the block */ 565d4f98a2SYan Zheng struct extent_buffer *eb; 575d4f98a2SYan Zheng /* level of tree block */ 585d4f98a2SYan Zheng unsigned int level:8; 595d4f98a2SYan Zheng /* 1 if the block is root of old snapshot */ 605d4f98a2SYan Zheng unsigned int old_root:1; 615d4f98a2SYan Zheng /* 1 if no child blocks in the cache */ 625d4f98a2SYan Zheng unsigned int lowest:1; 635d4f98a2SYan Zheng /* is the extent buffer locked */ 645d4f98a2SYan Zheng unsigned int locked:1; 655d4f98a2SYan Zheng /* has the block been processed */ 665d4f98a2SYan Zheng unsigned int processed:1; 675d4f98a2SYan Zheng /* have backrefs of this block been checked */ 685d4f98a2SYan Zheng unsigned int checked:1; 695d4f98a2SYan Zheng }; 705d4f98a2SYan Zheng 715d4f98a2SYan Zheng /* 725d4f98a2SYan Zheng * present a block pointer in the backref cache 735d4f98a2SYan Zheng */ 745d4f98a2SYan Zheng struct backref_edge { 755d4f98a2SYan Zheng struct list_head list[2]; 765d4f98a2SYan Zheng struct backref_node *node[2]; 775d4f98a2SYan Zheng u64 blockptr; 785d4f98a2SYan Zheng }; 795d4f98a2SYan Zheng 805d4f98a2SYan Zheng #define LOWER 0 815d4f98a2SYan Zheng #define UPPER 1 825d4f98a2SYan Zheng 835d4f98a2SYan Zheng struct backref_cache { 845d4f98a2SYan Zheng /* red black tree of all backref nodes in the cache */ 855d4f98a2SYan Zheng struct rb_root rb_root; 865d4f98a2SYan Zheng /* list of backref nodes with no child block in the cache */ 875d4f98a2SYan Zheng struct list_head pending[BTRFS_MAX_LEVEL]; 885d4f98a2SYan Zheng spinlock_t lock; 895d4f98a2SYan Zheng }; 905d4f98a2SYan Zheng 915d4f98a2SYan Zheng /* 925d4f98a2SYan Zheng * map address of tree root to tree 935d4f98a2SYan Zheng */ 945d4f98a2SYan Zheng struct mapping_node { 955d4f98a2SYan Zheng struct rb_node rb_node; 965d4f98a2SYan Zheng u64 bytenr; 975d4f98a2SYan Zheng void *data; 985d4f98a2SYan Zheng }; 995d4f98a2SYan Zheng 1005d4f98a2SYan Zheng struct mapping_tree { 1015d4f98a2SYan Zheng struct rb_root rb_root; 1025d4f98a2SYan Zheng spinlock_t lock; 1035d4f98a2SYan Zheng }; 1045d4f98a2SYan Zheng 1055d4f98a2SYan Zheng /* 1065d4f98a2SYan Zheng * present a tree block to process 1075d4f98a2SYan Zheng */ 1085d4f98a2SYan Zheng struct tree_block { 1095d4f98a2SYan Zheng struct rb_node rb_node; 1105d4f98a2SYan Zheng u64 bytenr; 1115d4f98a2SYan Zheng struct btrfs_key key; 1125d4f98a2SYan Zheng unsigned int level:8; 1135d4f98a2SYan Zheng unsigned int key_ready:1; 1145d4f98a2SYan Zheng }; 1155d4f98a2SYan Zheng 1165d4f98a2SYan Zheng /* inode vector */ 1175d4f98a2SYan Zheng #define INODEVEC_SIZE 16 1185d4f98a2SYan Zheng 1195d4f98a2SYan Zheng struct inodevec { 1205d4f98a2SYan Zheng struct list_head list; 1215d4f98a2SYan Zheng struct inode *inode[INODEVEC_SIZE]; 1225d4f98a2SYan Zheng int nr; 1235d4f98a2SYan Zheng }; 1245d4f98a2SYan Zheng 1250257bb82SYan, Zheng #define MAX_EXTENTS 128 1260257bb82SYan, Zheng 1270257bb82SYan, Zheng struct file_extent_cluster { 1280257bb82SYan, Zheng u64 start; 1290257bb82SYan, Zheng u64 end; 1300257bb82SYan, Zheng u64 boundary[MAX_EXTENTS]; 1310257bb82SYan, Zheng unsigned int nr; 1320257bb82SYan, Zheng }; 1330257bb82SYan, Zheng 1345d4f98a2SYan Zheng struct reloc_control { 1355d4f98a2SYan Zheng /* block group to relocate */ 1365d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group; 1375d4f98a2SYan Zheng /* extent tree */ 1385d4f98a2SYan Zheng struct btrfs_root *extent_root; 1395d4f98a2SYan Zheng /* inode for moving data */ 1405d4f98a2SYan Zheng struct inode *data_inode; 1415d4f98a2SYan Zheng struct btrfs_workers workers; 1425d4f98a2SYan Zheng /* tree blocks have been processed */ 1435d4f98a2SYan Zheng struct extent_io_tree processed_blocks; 1445d4f98a2SYan Zheng /* map start of tree root to corresponding reloc tree */ 1455d4f98a2SYan Zheng struct mapping_tree reloc_root_tree; 1465d4f98a2SYan Zheng /* list of reloc trees */ 1475d4f98a2SYan Zheng struct list_head reloc_roots; 1485d4f98a2SYan Zheng u64 search_start; 1495d4f98a2SYan Zheng u64 extents_found; 1505d4f98a2SYan Zheng u64 extents_skipped; 1515d4f98a2SYan Zheng int stage; 1525d4f98a2SYan Zheng int create_reloc_root; 1535d4f98a2SYan Zheng unsigned int found_file_extent:1; 1545d4f98a2SYan Zheng unsigned int found_old_snapshot:1; 1555d4f98a2SYan Zheng }; 1565d4f98a2SYan Zheng 1575d4f98a2SYan Zheng /* stages of data relocation */ 1585d4f98a2SYan Zheng #define MOVE_DATA_EXTENTS 0 1595d4f98a2SYan Zheng #define UPDATE_DATA_PTRS 1 1605d4f98a2SYan Zheng 1615d4f98a2SYan Zheng /* 1625d4f98a2SYan Zheng * merge reloc tree to corresponding fs tree in worker threads 1635d4f98a2SYan Zheng */ 1645d4f98a2SYan Zheng struct async_merge { 1655d4f98a2SYan Zheng struct btrfs_work work; 1665d4f98a2SYan Zheng struct reloc_control *rc; 1675d4f98a2SYan Zheng struct btrfs_root *root; 1685d4f98a2SYan Zheng struct completion *done; 1695d4f98a2SYan Zheng atomic_t *num_pending; 1705d4f98a2SYan Zheng }; 1715d4f98a2SYan Zheng 1725d4f98a2SYan Zheng static void mapping_tree_init(struct mapping_tree *tree) 1735d4f98a2SYan Zheng { 1746bef4d31SEric Paris tree->rb_root = RB_ROOT; 1755d4f98a2SYan Zheng spin_lock_init(&tree->lock); 1765d4f98a2SYan Zheng } 1775d4f98a2SYan Zheng 1785d4f98a2SYan Zheng static void backref_cache_init(struct backref_cache *cache) 1795d4f98a2SYan Zheng { 1805d4f98a2SYan Zheng int i; 1816bef4d31SEric Paris cache->rb_root = RB_ROOT; 1825d4f98a2SYan Zheng for (i = 0; i < BTRFS_MAX_LEVEL; i++) 1835d4f98a2SYan Zheng INIT_LIST_HEAD(&cache->pending[i]); 1845d4f98a2SYan Zheng spin_lock_init(&cache->lock); 1855d4f98a2SYan Zheng } 1865d4f98a2SYan Zheng 1875d4f98a2SYan Zheng static void backref_node_init(struct backref_node *node) 1885d4f98a2SYan Zheng { 1895d4f98a2SYan Zheng memset(node, 0, sizeof(*node)); 1905d4f98a2SYan Zheng INIT_LIST_HEAD(&node->upper); 1915d4f98a2SYan Zheng INIT_LIST_HEAD(&node->lower); 1925d4f98a2SYan Zheng RB_CLEAR_NODE(&node->rb_node); 1935d4f98a2SYan Zheng } 1945d4f98a2SYan Zheng 1955d4f98a2SYan Zheng static struct rb_node *tree_insert(struct rb_root *root, u64 bytenr, 1965d4f98a2SYan Zheng struct rb_node *node) 1975d4f98a2SYan Zheng { 1985d4f98a2SYan Zheng struct rb_node **p = &root->rb_node; 1995d4f98a2SYan Zheng struct rb_node *parent = NULL; 2005d4f98a2SYan Zheng struct tree_entry *entry; 2015d4f98a2SYan Zheng 2025d4f98a2SYan Zheng while (*p) { 2035d4f98a2SYan Zheng parent = *p; 2045d4f98a2SYan Zheng entry = rb_entry(parent, struct tree_entry, rb_node); 2055d4f98a2SYan Zheng 2065d4f98a2SYan Zheng if (bytenr < entry->bytenr) 2075d4f98a2SYan Zheng p = &(*p)->rb_left; 2085d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 2095d4f98a2SYan Zheng p = &(*p)->rb_right; 2105d4f98a2SYan Zheng else 2115d4f98a2SYan Zheng return parent; 2125d4f98a2SYan Zheng } 2135d4f98a2SYan Zheng 2145d4f98a2SYan Zheng rb_link_node(node, parent, p); 2155d4f98a2SYan Zheng rb_insert_color(node, root); 2165d4f98a2SYan Zheng return NULL; 2175d4f98a2SYan Zheng } 2185d4f98a2SYan Zheng 2195d4f98a2SYan Zheng static struct rb_node *tree_search(struct rb_root *root, u64 bytenr) 2205d4f98a2SYan Zheng { 2215d4f98a2SYan Zheng struct rb_node *n = root->rb_node; 2225d4f98a2SYan Zheng struct tree_entry *entry; 2235d4f98a2SYan Zheng 2245d4f98a2SYan Zheng while (n) { 2255d4f98a2SYan Zheng entry = rb_entry(n, struct tree_entry, rb_node); 2265d4f98a2SYan Zheng 2275d4f98a2SYan Zheng if (bytenr < entry->bytenr) 2285d4f98a2SYan Zheng n = n->rb_left; 2295d4f98a2SYan Zheng else if (bytenr > entry->bytenr) 2305d4f98a2SYan Zheng n = n->rb_right; 2315d4f98a2SYan Zheng else 2325d4f98a2SYan Zheng return n; 2335d4f98a2SYan Zheng } 2345d4f98a2SYan Zheng return NULL; 2355d4f98a2SYan Zheng } 2365d4f98a2SYan Zheng 2375d4f98a2SYan Zheng /* 2385d4f98a2SYan Zheng * walk up backref nodes until reach node presents tree root 2395d4f98a2SYan Zheng */ 2405d4f98a2SYan Zheng static struct backref_node *walk_up_backref(struct backref_node *node, 2415d4f98a2SYan Zheng struct backref_edge *edges[], 2425d4f98a2SYan Zheng int *index) 2435d4f98a2SYan Zheng { 2445d4f98a2SYan Zheng struct backref_edge *edge; 2455d4f98a2SYan Zheng int idx = *index; 2465d4f98a2SYan Zheng 2475d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 2485d4f98a2SYan Zheng edge = list_entry(node->upper.next, 2495d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 2505d4f98a2SYan Zheng edges[idx++] = edge; 2515d4f98a2SYan Zheng node = edge->node[UPPER]; 2525d4f98a2SYan Zheng } 2535d4f98a2SYan Zheng *index = idx; 2545d4f98a2SYan Zheng return node; 2555d4f98a2SYan Zheng } 2565d4f98a2SYan Zheng 2575d4f98a2SYan Zheng /* 2585d4f98a2SYan Zheng * walk down backref nodes to find start of next reference path 2595d4f98a2SYan Zheng */ 2605d4f98a2SYan Zheng static struct backref_node *walk_down_backref(struct backref_edge *edges[], 2615d4f98a2SYan Zheng int *index) 2625d4f98a2SYan Zheng { 2635d4f98a2SYan Zheng struct backref_edge *edge; 2645d4f98a2SYan Zheng struct backref_node *lower; 2655d4f98a2SYan Zheng int idx = *index; 2665d4f98a2SYan Zheng 2675d4f98a2SYan Zheng while (idx > 0) { 2685d4f98a2SYan Zheng edge = edges[idx - 1]; 2695d4f98a2SYan Zheng lower = edge->node[LOWER]; 2705d4f98a2SYan Zheng if (list_is_last(&edge->list[LOWER], &lower->upper)) { 2715d4f98a2SYan Zheng idx--; 2725d4f98a2SYan Zheng continue; 2735d4f98a2SYan Zheng } 2745d4f98a2SYan Zheng edge = list_entry(edge->list[LOWER].next, 2755d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 2765d4f98a2SYan Zheng edges[idx - 1] = edge; 2775d4f98a2SYan Zheng *index = idx; 2785d4f98a2SYan Zheng return edge->node[UPPER]; 2795d4f98a2SYan Zheng } 2805d4f98a2SYan Zheng *index = 0; 2815d4f98a2SYan Zheng return NULL; 2825d4f98a2SYan Zheng } 2835d4f98a2SYan Zheng 2845d4f98a2SYan Zheng static void drop_node_buffer(struct backref_node *node) 2855d4f98a2SYan Zheng { 2865d4f98a2SYan Zheng if (node->eb) { 2875d4f98a2SYan Zheng if (node->locked) { 2885d4f98a2SYan Zheng btrfs_tree_unlock(node->eb); 2895d4f98a2SYan Zheng node->locked = 0; 2905d4f98a2SYan Zheng } 2915d4f98a2SYan Zheng free_extent_buffer(node->eb); 2925d4f98a2SYan Zheng node->eb = NULL; 2935d4f98a2SYan Zheng } 2945d4f98a2SYan Zheng } 2955d4f98a2SYan Zheng 2965d4f98a2SYan Zheng static void drop_backref_node(struct backref_cache *tree, 2975d4f98a2SYan Zheng struct backref_node *node) 2985d4f98a2SYan Zheng { 2995d4f98a2SYan Zheng BUG_ON(!node->lowest); 3005d4f98a2SYan Zheng BUG_ON(!list_empty(&node->upper)); 3015d4f98a2SYan Zheng 3025d4f98a2SYan Zheng drop_node_buffer(node); 3035d4f98a2SYan Zheng list_del(&node->lower); 3045d4f98a2SYan Zheng 3055d4f98a2SYan Zheng rb_erase(&node->rb_node, &tree->rb_root); 3065d4f98a2SYan Zheng kfree(node); 3075d4f98a2SYan Zheng } 3085d4f98a2SYan Zheng 3095d4f98a2SYan Zheng /* 3105d4f98a2SYan Zheng * remove a backref node from the backref cache 3115d4f98a2SYan Zheng */ 3125d4f98a2SYan Zheng static void remove_backref_node(struct backref_cache *cache, 3135d4f98a2SYan Zheng struct backref_node *node) 3145d4f98a2SYan Zheng { 3155d4f98a2SYan Zheng struct backref_node *upper; 3165d4f98a2SYan Zheng struct backref_edge *edge; 3175d4f98a2SYan Zheng 3185d4f98a2SYan Zheng if (!node) 3195d4f98a2SYan Zheng return; 3205d4f98a2SYan Zheng 3215d4f98a2SYan Zheng BUG_ON(!node->lowest); 3225d4f98a2SYan Zheng while (!list_empty(&node->upper)) { 3235d4f98a2SYan Zheng edge = list_entry(node->upper.next, struct backref_edge, 3245d4f98a2SYan Zheng list[LOWER]); 3255d4f98a2SYan Zheng upper = edge->node[UPPER]; 3265d4f98a2SYan Zheng list_del(&edge->list[LOWER]); 3275d4f98a2SYan Zheng list_del(&edge->list[UPPER]); 3285d4f98a2SYan Zheng kfree(edge); 3295d4f98a2SYan Zheng /* 3305d4f98a2SYan Zheng * add the node to pending list if no other 3315d4f98a2SYan Zheng * child block cached. 3325d4f98a2SYan Zheng */ 3335d4f98a2SYan Zheng if (list_empty(&upper->lower)) { 3345d4f98a2SYan Zheng list_add_tail(&upper->lower, 3355d4f98a2SYan Zheng &cache->pending[upper->level]); 3365d4f98a2SYan Zheng upper->lowest = 1; 3375d4f98a2SYan Zheng } 3385d4f98a2SYan Zheng } 3395d4f98a2SYan Zheng drop_backref_node(cache, node); 3405d4f98a2SYan Zheng } 3415d4f98a2SYan Zheng 3425d4f98a2SYan Zheng /* 3435d4f98a2SYan Zheng * find reloc tree by address of tree root 3445d4f98a2SYan Zheng */ 3455d4f98a2SYan Zheng static struct btrfs_root *find_reloc_root(struct reloc_control *rc, 3465d4f98a2SYan Zheng u64 bytenr) 3475d4f98a2SYan Zheng { 3485d4f98a2SYan Zheng struct rb_node *rb_node; 3495d4f98a2SYan Zheng struct mapping_node *node; 3505d4f98a2SYan Zheng struct btrfs_root *root = NULL; 3515d4f98a2SYan Zheng 3525d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 3535d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, bytenr); 3545d4f98a2SYan Zheng if (rb_node) { 3555d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 3565d4f98a2SYan Zheng root = (struct btrfs_root *)node->data; 3575d4f98a2SYan Zheng } 3585d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 3595d4f98a2SYan Zheng return root; 3605d4f98a2SYan Zheng } 3615d4f98a2SYan Zheng 3625d4f98a2SYan Zheng static int is_cowonly_root(u64 root_objectid) 3635d4f98a2SYan Zheng { 3645d4f98a2SYan Zheng if (root_objectid == BTRFS_ROOT_TREE_OBJECTID || 3655d4f98a2SYan Zheng root_objectid == BTRFS_EXTENT_TREE_OBJECTID || 3665d4f98a2SYan Zheng root_objectid == BTRFS_CHUNK_TREE_OBJECTID || 3675d4f98a2SYan Zheng root_objectid == BTRFS_DEV_TREE_OBJECTID || 3685d4f98a2SYan Zheng root_objectid == BTRFS_TREE_LOG_OBJECTID || 3695d4f98a2SYan Zheng root_objectid == BTRFS_CSUM_TREE_OBJECTID) 3705d4f98a2SYan Zheng return 1; 3715d4f98a2SYan Zheng return 0; 3725d4f98a2SYan Zheng } 3735d4f98a2SYan Zheng 3745d4f98a2SYan Zheng static struct btrfs_root *read_fs_root(struct btrfs_fs_info *fs_info, 3755d4f98a2SYan Zheng u64 root_objectid) 3765d4f98a2SYan Zheng { 3775d4f98a2SYan Zheng struct btrfs_key key; 3785d4f98a2SYan Zheng 3795d4f98a2SYan Zheng key.objectid = root_objectid; 3805d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 3815d4f98a2SYan Zheng if (is_cowonly_root(root_objectid)) 3825d4f98a2SYan Zheng key.offset = 0; 3835d4f98a2SYan Zheng else 3845d4f98a2SYan Zheng key.offset = (u64)-1; 3855d4f98a2SYan Zheng 3865d4f98a2SYan Zheng return btrfs_read_fs_root_no_name(fs_info, &key); 3875d4f98a2SYan Zheng } 3885d4f98a2SYan Zheng 3895d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 3905d4f98a2SYan Zheng static noinline_for_stack 3915d4f98a2SYan Zheng struct btrfs_root *find_tree_root(struct reloc_control *rc, 3925d4f98a2SYan Zheng struct extent_buffer *leaf, 3935d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0) 3945d4f98a2SYan Zheng { 3955d4f98a2SYan Zheng struct btrfs_root *root; 3965d4f98a2SYan Zheng u64 root_objectid = btrfs_ref_root_v0(leaf, ref0); 3975d4f98a2SYan Zheng u64 generation = btrfs_ref_generation_v0(leaf, ref0); 3985d4f98a2SYan Zheng 3995d4f98a2SYan Zheng BUG_ON(root_objectid == BTRFS_TREE_RELOC_OBJECTID); 4005d4f98a2SYan Zheng 4015d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, root_objectid); 4025d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 4035d4f98a2SYan Zheng 4045d4f98a2SYan Zheng if (root->ref_cows && 4055d4f98a2SYan Zheng generation != btrfs_root_generation(&root->root_item)) 4065d4f98a2SYan Zheng return NULL; 4075d4f98a2SYan Zheng 4085d4f98a2SYan Zheng return root; 4095d4f98a2SYan Zheng } 4105d4f98a2SYan Zheng #endif 4115d4f98a2SYan Zheng 4125d4f98a2SYan Zheng static noinline_for_stack 4135d4f98a2SYan Zheng int find_inline_backref(struct extent_buffer *leaf, int slot, 4145d4f98a2SYan Zheng unsigned long *ptr, unsigned long *end) 4155d4f98a2SYan Zheng { 4165d4f98a2SYan Zheng struct btrfs_extent_item *ei; 4175d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 4185d4f98a2SYan Zheng u32 item_size; 4195d4f98a2SYan Zheng 4205d4f98a2SYan Zheng item_size = btrfs_item_size_nr(leaf, slot); 4215d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 4225d4f98a2SYan Zheng if (item_size < sizeof(*ei)) { 4235d4f98a2SYan Zheng WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 4245d4f98a2SYan Zheng return 1; 4255d4f98a2SYan Zheng } 4265d4f98a2SYan Zheng #endif 4275d4f98a2SYan Zheng ei = btrfs_item_ptr(leaf, slot, struct btrfs_extent_item); 4285d4f98a2SYan Zheng WARN_ON(!(btrfs_extent_flags(leaf, ei) & 4295d4f98a2SYan Zheng BTRFS_EXTENT_FLAG_TREE_BLOCK)); 4305d4f98a2SYan Zheng 4315d4f98a2SYan Zheng if (item_size <= sizeof(*ei) + sizeof(*bi)) { 4325d4f98a2SYan Zheng WARN_ON(item_size < sizeof(*ei) + sizeof(*bi)); 4335d4f98a2SYan Zheng return 1; 4345d4f98a2SYan Zheng } 4355d4f98a2SYan Zheng 4365d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 4375d4f98a2SYan Zheng *ptr = (unsigned long)(bi + 1); 4385d4f98a2SYan Zheng *end = (unsigned long)ei + item_size; 4395d4f98a2SYan Zheng return 0; 4405d4f98a2SYan Zheng } 4415d4f98a2SYan Zheng 4425d4f98a2SYan Zheng /* 4435d4f98a2SYan Zheng * build backref tree for a given tree block. root of the backref tree 4445d4f98a2SYan Zheng * corresponds the tree block, leaves of the backref tree correspond 4455d4f98a2SYan Zheng * roots of b-trees that reference the tree block. 4465d4f98a2SYan Zheng * 4475d4f98a2SYan Zheng * the basic idea of this function is check backrefs of a given block 4485d4f98a2SYan Zheng * to find upper level blocks that refernece the block, and then check 4495d4f98a2SYan Zheng * bakcrefs of these upper level blocks recursively. the recursion stop 4505d4f98a2SYan Zheng * when tree root is reached or backrefs for the block is cached. 4515d4f98a2SYan Zheng * 4525d4f98a2SYan Zheng * NOTE: if we find backrefs for a block are cached, we know backrefs 4535d4f98a2SYan Zheng * for all upper level blocks that directly/indirectly reference the 4545d4f98a2SYan Zheng * block are also cached. 4555d4f98a2SYan Zheng */ 4565d4f98a2SYan Zheng static struct backref_node *build_backref_tree(struct reloc_control *rc, 4575d4f98a2SYan Zheng struct backref_cache *cache, 4585d4f98a2SYan Zheng struct btrfs_key *node_key, 4595d4f98a2SYan Zheng int level, u64 bytenr) 4605d4f98a2SYan Zheng { 4615d4f98a2SYan Zheng struct btrfs_path *path1; 4625d4f98a2SYan Zheng struct btrfs_path *path2; 4635d4f98a2SYan Zheng struct extent_buffer *eb; 4645d4f98a2SYan Zheng struct btrfs_root *root; 4655d4f98a2SYan Zheng struct backref_node *cur; 4665d4f98a2SYan Zheng struct backref_node *upper; 4675d4f98a2SYan Zheng struct backref_node *lower; 4685d4f98a2SYan Zheng struct backref_node *node = NULL; 4695d4f98a2SYan Zheng struct backref_node *exist = NULL; 4705d4f98a2SYan Zheng struct backref_edge *edge; 4715d4f98a2SYan Zheng struct rb_node *rb_node; 4725d4f98a2SYan Zheng struct btrfs_key key; 4735d4f98a2SYan Zheng unsigned long end; 4745d4f98a2SYan Zheng unsigned long ptr; 4755d4f98a2SYan Zheng LIST_HEAD(list); 4765d4f98a2SYan Zheng int ret; 4775d4f98a2SYan Zheng int err = 0; 4785d4f98a2SYan Zheng 4795d4f98a2SYan Zheng path1 = btrfs_alloc_path(); 4805d4f98a2SYan Zheng path2 = btrfs_alloc_path(); 4815d4f98a2SYan Zheng if (!path1 || !path2) { 4825d4f98a2SYan Zheng err = -ENOMEM; 4835d4f98a2SYan Zheng goto out; 4845d4f98a2SYan Zheng } 4855d4f98a2SYan Zheng 4865d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 4875d4f98a2SYan Zheng if (!node) { 4885d4f98a2SYan Zheng err = -ENOMEM; 4895d4f98a2SYan Zheng goto out; 4905d4f98a2SYan Zheng } 4915d4f98a2SYan Zheng 4925d4f98a2SYan Zheng backref_node_init(node); 4935d4f98a2SYan Zheng node->bytenr = bytenr; 4945d4f98a2SYan Zheng node->owner = 0; 4955d4f98a2SYan Zheng node->level = level; 4965d4f98a2SYan Zheng node->lowest = 1; 4975d4f98a2SYan Zheng cur = node; 4985d4f98a2SYan Zheng again: 4995d4f98a2SYan Zheng end = 0; 5005d4f98a2SYan Zheng ptr = 0; 5015d4f98a2SYan Zheng key.objectid = cur->bytenr; 5025d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 5035d4f98a2SYan Zheng key.offset = (u64)-1; 5045d4f98a2SYan Zheng 5055d4f98a2SYan Zheng path1->search_commit_root = 1; 5065d4f98a2SYan Zheng path1->skip_locking = 1; 5075d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path1, 5085d4f98a2SYan Zheng 0, 0); 5095d4f98a2SYan Zheng if (ret < 0) { 5105d4f98a2SYan Zheng err = ret; 5115d4f98a2SYan Zheng goto out; 5125d4f98a2SYan Zheng } 5135d4f98a2SYan Zheng BUG_ON(!ret || !path1->slots[0]); 5145d4f98a2SYan Zheng 5155d4f98a2SYan Zheng path1->slots[0]--; 5165d4f98a2SYan Zheng 5175d4f98a2SYan Zheng WARN_ON(cur->checked); 5185d4f98a2SYan Zheng if (!list_empty(&cur->upper)) { 5195d4f98a2SYan Zheng /* 5205d4f98a2SYan Zheng * the backref was added previously when processsing 5215d4f98a2SYan Zheng * backref of type BTRFS_TREE_BLOCK_REF_KEY 5225d4f98a2SYan Zheng */ 5235d4f98a2SYan Zheng BUG_ON(!list_is_singular(&cur->upper)); 5245d4f98a2SYan Zheng edge = list_entry(cur->upper.next, struct backref_edge, 5255d4f98a2SYan Zheng list[LOWER]); 5265d4f98a2SYan Zheng BUG_ON(!list_empty(&edge->list[UPPER])); 5275d4f98a2SYan Zheng exist = edge->node[UPPER]; 5285d4f98a2SYan Zheng /* 5295d4f98a2SYan Zheng * add the upper level block to pending list if we need 5305d4f98a2SYan Zheng * check its backrefs 5315d4f98a2SYan Zheng */ 5325d4f98a2SYan Zheng if (!exist->checked) 5335d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 5345d4f98a2SYan Zheng } else { 5355d4f98a2SYan Zheng exist = NULL; 5365d4f98a2SYan Zheng } 5375d4f98a2SYan Zheng 5385d4f98a2SYan Zheng while (1) { 5395d4f98a2SYan Zheng cond_resched(); 5405d4f98a2SYan Zheng eb = path1->nodes[0]; 5415d4f98a2SYan Zheng 5425d4f98a2SYan Zheng if (ptr >= end) { 5435d4f98a2SYan Zheng if (path1->slots[0] >= btrfs_header_nritems(eb)) { 5445d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path1); 5455d4f98a2SYan Zheng if (ret < 0) { 5465d4f98a2SYan Zheng err = ret; 5475d4f98a2SYan Zheng goto out; 5485d4f98a2SYan Zheng } 5495d4f98a2SYan Zheng if (ret > 0) 5505d4f98a2SYan Zheng break; 5515d4f98a2SYan Zheng eb = path1->nodes[0]; 5525d4f98a2SYan Zheng } 5535d4f98a2SYan Zheng 5545d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path1->slots[0]); 5555d4f98a2SYan Zheng if (key.objectid != cur->bytenr) { 5565d4f98a2SYan Zheng WARN_ON(exist); 5575d4f98a2SYan Zheng break; 5585d4f98a2SYan Zheng } 5595d4f98a2SYan Zheng 5605d4f98a2SYan Zheng if (key.type == BTRFS_EXTENT_ITEM_KEY) { 5615d4f98a2SYan Zheng ret = find_inline_backref(eb, path1->slots[0], 5625d4f98a2SYan Zheng &ptr, &end); 5635d4f98a2SYan Zheng if (ret) 5645d4f98a2SYan Zheng goto next; 5655d4f98a2SYan Zheng } 5665d4f98a2SYan Zheng } 5675d4f98a2SYan Zheng 5685d4f98a2SYan Zheng if (ptr < end) { 5695d4f98a2SYan Zheng /* update key for inline back ref */ 5705d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 5715d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 5725d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 5735d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 5745d4f98a2SYan Zheng WARN_ON(key.type != BTRFS_TREE_BLOCK_REF_KEY && 5755d4f98a2SYan Zheng key.type != BTRFS_SHARED_BLOCK_REF_KEY); 5765d4f98a2SYan Zheng } 5775d4f98a2SYan Zheng 5785d4f98a2SYan Zheng if (exist && 5795d4f98a2SYan Zheng ((key.type == BTRFS_TREE_BLOCK_REF_KEY && 5805d4f98a2SYan Zheng exist->owner == key.offset) || 5815d4f98a2SYan Zheng (key.type == BTRFS_SHARED_BLOCK_REF_KEY && 5825d4f98a2SYan Zheng exist->bytenr == key.offset))) { 5835d4f98a2SYan Zheng exist = NULL; 5845d4f98a2SYan Zheng goto next; 5855d4f98a2SYan Zheng } 5865d4f98a2SYan Zheng 5875d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 5885d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY || 5895d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 5905d4f98a2SYan Zheng if (key.objectid == key.offset && 5915d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 5925d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 5935d4f98a2SYan Zheng ref0 = btrfs_item_ptr(eb, path1->slots[0], 5945d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 5955d4f98a2SYan Zheng root = find_tree_root(rc, eb, ref0); 5965d4f98a2SYan Zheng if (root) 5975d4f98a2SYan Zheng cur->root = root; 5985d4f98a2SYan Zheng else 5995d4f98a2SYan Zheng cur->old_root = 1; 6005d4f98a2SYan Zheng break; 6015d4f98a2SYan Zheng } 6025d4f98a2SYan Zheng #else 6035d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 6045d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_BLOCK_REF_KEY) { 6055d4f98a2SYan Zheng #endif 6065d4f98a2SYan Zheng if (key.objectid == key.offset) { 6075d4f98a2SYan Zheng /* 6085d4f98a2SYan Zheng * only root blocks of reloc trees use 6095d4f98a2SYan Zheng * backref of this type. 6105d4f98a2SYan Zheng */ 6115d4f98a2SYan Zheng root = find_reloc_root(rc, cur->bytenr); 6125d4f98a2SYan Zheng BUG_ON(!root); 6135d4f98a2SYan Zheng cur->root = root; 6145d4f98a2SYan Zheng break; 6155d4f98a2SYan Zheng } 6165d4f98a2SYan Zheng 6175d4f98a2SYan Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 6185d4f98a2SYan Zheng if (!edge) { 6195d4f98a2SYan Zheng err = -ENOMEM; 6205d4f98a2SYan Zheng goto out; 6215d4f98a2SYan Zheng } 6225d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, key.offset); 6235d4f98a2SYan Zheng if (!rb_node) { 6245d4f98a2SYan Zheng upper = kmalloc(sizeof(*upper), GFP_NOFS); 6255d4f98a2SYan Zheng if (!upper) { 6265d4f98a2SYan Zheng kfree(edge); 6275d4f98a2SYan Zheng err = -ENOMEM; 6285d4f98a2SYan Zheng goto out; 6295d4f98a2SYan Zheng } 6305d4f98a2SYan Zheng backref_node_init(upper); 6315d4f98a2SYan Zheng upper->bytenr = key.offset; 6325d4f98a2SYan Zheng upper->owner = 0; 6335d4f98a2SYan Zheng upper->level = cur->level + 1; 6345d4f98a2SYan Zheng /* 6355d4f98a2SYan Zheng * backrefs for the upper level block isn't 6365d4f98a2SYan Zheng * cached, add the block to pending list 6375d4f98a2SYan Zheng */ 6385d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 6395d4f98a2SYan Zheng } else { 6405d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 6415d4f98a2SYan Zheng rb_node); 6425d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 6435d4f98a2SYan Zheng } 6445d4f98a2SYan Zheng list_add(&edge->list[LOWER], &cur->upper); 6455d4f98a2SYan Zheng edge->node[UPPER] = upper; 6465d4f98a2SYan Zheng edge->node[LOWER] = cur; 6475d4f98a2SYan Zheng 6485d4f98a2SYan Zheng goto next; 6495d4f98a2SYan Zheng } else if (key.type != BTRFS_TREE_BLOCK_REF_KEY) { 6505d4f98a2SYan Zheng goto next; 6515d4f98a2SYan Zheng } 6525d4f98a2SYan Zheng 6535d4f98a2SYan Zheng /* key.type == BTRFS_TREE_BLOCK_REF_KEY */ 6545d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, key.offset); 6555d4f98a2SYan Zheng if (IS_ERR(root)) { 6565d4f98a2SYan Zheng err = PTR_ERR(root); 6575d4f98a2SYan Zheng goto out; 6585d4f98a2SYan Zheng } 6595d4f98a2SYan Zheng 6605d4f98a2SYan Zheng if (btrfs_root_level(&root->root_item) == cur->level) { 6615d4f98a2SYan Zheng /* tree root */ 6625d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 6635d4f98a2SYan Zheng cur->bytenr); 6645d4f98a2SYan Zheng cur->root = root; 6655d4f98a2SYan Zheng break; 6665d4f98a2SYan Zheng } 6675d4f98a2SYan Zheng 6685d4f98a2SYan Zheng level = cur->level + 1; 6695d4f98a2SYan Zheng 6705d4f98a2SYan Zheng /* 6715d4f98a2SYan Zheng * searching the tree to find upper level blocks 6725d4f98a2SYan Zheng * reference the block. 6735d4f98a2SYan Zheng */ 6745d4f98a2SYan Zheng path2->search_commit_root = 1; 6755d4f98a2SYan Zheng path2->skip_locking = 1; 6765d4f98a2SYan Zheng path2->lowest_level = level; 6775d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, node_key, path2, 0, 0); 6785d4f98a2SYan Zheng path2->lowest_level = 0; 6795d4f98a2SYan Zheng if (ret < 0) { 6805d4f98a2SYan Zheng err = ret; 6815d4f98a2SYan Zheng goto out; 6825d4f98a2SYan Zheng } 68333c66f43SYan Zheng if (ret > 0 && path2->slots[level] > 0) 68433c66f43SYan Zheng path2->slots[level]--; 6855d4f98a2SYan Zheng 6865d4f98a2SYan Zheng eb = path2->nodes[level]; 6875d4f98a2SYan Zheng WARN_ON(btrfs_node_blockptr(eb, path2->slots[level]) != 6885d4f98a2SYan Zheng cur->bytenr); 6895d4f98a2SYan Zheng 6905d4f98a2SYan Zheng lower = cur; 6915d4f98a2SYan Zheng for (; level < BTRFS_MAX_LEVEL; level++) { 6925d4f98a2SYan Zheng if (!path2->nodes[level]) { 6935d4f98a2SYan Zheng BUG_ON(btrfs_root_bytenr(&root->root_item) != 6945d4f98a2SYan Zheng lower->bytenr); 6955d4f98a2SYan Zheng lower->root = root; 6965d4f98a2SYan Zheng break; 6975d4f98a2SYan Zheng } 6985d4f98a2SYan Zheng 6995d4f98a2SYan Zheng edge = kzalloc(sizeof(*edge), GFP_NOFS); 7005d4f98a2SYan Zheng if (!edge) { 7015d4f98a2SYan Zheng err = -ENOMEM; 7025d4f98a2SYan Zheng goto out; 7035d4f98a2SYan Zheng } 7045d4f98a2SYan Zheng 7055d4f98a2SYan Zheng eb = path2->nodes[level]; 7065d4f98a2SYan Zheng rb_node = tree_search(&cache->rb_root, eb->start); 7075d4f98a2SYan Zheng if (!rb_node) { 7085d4f98a2SYan Zheng upper = kmalloc(sizeof(*upper), GFP_NOFS); 7095d4f98a2SYan Zheng if (!upper) { 7105d4f98a2SYan Zheng kfree(edge); 7115d4f98a2SYan Zheng err = -ENOMEM; 7125d4f98a2SYan Zheng goto out; 7135d4f98a2SYan Zheng } 7145d4f98a2SYan Zheng backref_node_init(upper); 7155d4f98a2SYan Zheng upper->bytenr = eb->start; 7165d4f98a2SYan Zheng upper->owner = btrfs_header_owner(eb); 7175d4f98a2SYan Zheng upper->level = lower->level + 1; 7185d4f98a2SYan Zheng 7195d4f98a2SYan Zheng /* 7205d4f98a2SYan Zheng * if we know the block isn't shared 7215d4f98a2SYan Zheng * we can void checking its backrefs. 7225d4f98a2SYan Zheng */ 7235d4f98a2SYan Zheng if (btrfs_block_can_be_shared(root, eb)) 7245d4f98a2SYan Zheng upper->checked = 0; 7255d4f98a2SYan Zheng else 7265d4f98a2SYan Zheng upper->checked = 1; 7275d4f98a2SYan Zheng 7285d4f98a2SYan Zheng /* 7295d4f98a2SYan Zheng * add the block to pending list if we 7305d4f98a2SYan Zheng * need check its backrefs. only block 7315d4f98a2SYan Zheng * at 'cur->level + 1' is added to the 7325d4f98a2SYan Zheng * tail of pending list. this guarantees 7335d4f98a2SYan Zheng * we check backrefs from lower level 7345d4f98a2SYan Zheng * blocks to upper level blocks. 7355d4f98a2SYan Zheng */ 7365d4f98a2SYan Zheng if (!upper->checked && 7375d4f98a2SYan Zheng level == cur->level + 1) { 7385d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], 7395d4f98a2SYan Zheng &list); 7405d4f98a2SYan Zheng } else 7415d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 7425d4f98a2SYan Zheng } else { 7435d4f98a2SYan Zheng upper = rb_entry(rb_node, struct backref_node, 7445d4f98a2SYan Zheng rb_node); 7455d4f98a2SYan Zheng BUG_ON(!upper->checked); 7465d4f98a2SYan Zheng INIT_LIST_HEAD(&edge->list[UPPER]); 7475d4f98a2SYan Zheng } 7485d4f98a2SYan Zheng list_add_tail(&edge->list[LOWER], &lower->upper); 7495d4f98a2SYan Zheng edge->node[UPPER] = upper; 7505d4f98a2SYan Zheng edge->node[LOWER] = lower; 7515d4f98a2SYan Zheng 7525d4f98a2SYan Zheng if (rb_node) 7535d4f98a2SYan Zheng break; 7545d4f98a2SYan Zheng lower = upper; 7555d4f98a2SYan Zheng upper = NULL; 7565d4f98a2SYan Zheng } 7575d4f98a2SYan Zheng btrfs_release_path(root, path2); 7585d4f98a2SYan Zheng next: 7595d4f98a2SYan Zheng if (ptr < end) { 7605d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 7615d4f98a2SYan Zheng if (ptr >= end) { 7625d4f98a2SYan Zheng WARN_ON(ptr > end); 7635d4f98a2SYan Zheng ptr = 0; 7645d4f98a2SYan Zheng end = 0; 7655d4f98a2SYan Zheng } 7665d4f98a2SYan Zheng } 7675d4f98a2SYan Zheng if (ptr >= end) 7685d4f98a2SYan Zheng path1->slots[0]++; 7695d4f98a2SYan Zheng } 7705d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path1); 7715d4f98a2SYan Zheng 7725d4f98a2SYan Zheng cur->checked = 1; 7735d4f98a2SYan Zheng WARN_ON(exist); 7745d4f98a2SYan Zheng 7755d4f98a2SYan Zheng /* the pending list isn't empty, take the first block to process */ 7765d4f98a2SYan Zheng if (!list_empty(&list)) { 7775d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 7785d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 7795d4f98a2SYan Zheng cur = edge->node[UPPER]; 7805d4f98a2SYan Zheng goto again; 7815d4f98a2SYan Zheng } 7825d4f98a2SYan Zheng 7835d4f98a2SYan Zheng /* 7845d4f98a2SYan Zheng * everything goes well, connect backref nodes and insert backref nodes 7855d4f98a2SYan Zheng * into the cache. 7865d4f98a2SYan Zheng */ 7875d4f98a2SYan Zheng BUG_ON(!node->checked); 7885d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, node->bytenr, &node->rb_node); 7895d4f98a2SYan Zheng BUG_ON(rb_node); 7905d4f98a2SYan Zheng 7915d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) 7925d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 7935d4f98a2SYan Zheng 7945d4f98a2SYan Zheng while (!list_empty(&list)) { 7955d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, list[UPPER]); 7965d4f98a2SYan Zheng list_del_init(&edge->list[UPPER]); 7975d4f98a2SYan Zheng upper = edge->node[UPPER]; 7985d4f98a2SYan Zheng 7995d4f98a2SYan Zheng if (!RB_EMPTY_NODE(&upper->rb_node)) { 8005d4f98a2SYan Zheng if (upper->lowest) { 8015d4f98a2SYan Zheng list_del_init(&upper->lower); 8025d4f98a2SYan Zheng upper->lowest = 0; 8035d4f98a2SYan Zheng } 8045d4f98a2SYan Zheng 8055d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 8065d4f98a2SYan Zheng continue; 8075d4f98a2SYan Zheng } 8085d4f98a2SYan Zheng 8095d4f98a2SYan Zheng BUG_ON(!upper->checked); 8105d4f98a2SYan Zheng rb_node = tree_insert(&cache->rb_root, upper->bytenr, 8115d4f98a2SYan Zheng &upper->rb_node); 8125d4f98a2SYan Zheng BUG_ON(rb_node); 8135d4f98a2SYan Zheng 8145d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &upper->lower); 8155d4f98a2SYan Zheng 8165d4f98a2SYan Zheng list_for_each_entry(edge, &upper->upper, list[LOWER]) 8175d4f98a2SYan Zheng list_add_tail(&edge->list[UPPER], &list); 8185d4f98a2SYan Zheng } 8195d4f98a2SYan Zheng out: 8205d4f98a2SYan Zheng btrfs_free_path(path1); 8215d4f98a2SYan Zheng btrfs_free_path(path2); 8225d4f98a2SYan Zheng if (err) { 8235d4f98a2SYan Zheng INIT_LIST_HEAD(&list); 8245d4f98a2SYan Zheng upper = node; 8255d4f98a2SYan Zheng while (upper) { 8265d4f98a2SYan Zheng if (RB_EMPTY_NODE(&upper->rb_node)) { 8275d4f98a2SYan Zheng list_splice_tail(&upper->upper, &list); 8285d4f98a2SYan Zheng kfree(upper); 8295d4f98a2SYan Zheng } 8305d4f98a2SYan Zheng 8315d4f98a2SYan Zheng if (list_empty(&list)) 8325d4f98a2SYan Zheng break; 8335d4f98a2SYan Zheng 8345d4f98a2SYan Zheng edge = list_entry(list.next, struct backref_edge, 8355d4f98a2SYan Zheng list[LOWER]); 8365d4f98a2SYan Zheng upper = edge->node[UPPER]; 8375d4f98a2SYan Zheng kfree(edge); 8385d4f98a2SYan Zheng } 8395d4f98a2SYan Zheng return ERR_PTR(err); 8405d4f98a2SYan Zheng } 8415d4f98a2SYan Zheng return node; 8425d4f98a2SYan Zheng } 8435d4f98a2SYan Zheng 8445d4f98a2SYan Zheng /* 8455d4f98a2SYan Zheng * helper to add 'address of tree root -> reloc tree' mapping 8465d4f98a2SYan Zheng */ 8475d4f98a2SYan Zheng static int __add_reloc_root(struct btrfs_root *root) 8485d4f98a2SYan Zheng { 8495d4f98a2SYan Zheng struct rb_node *rb_node; 8505d4f98a2SYan Zheng struct mapping_node *node; 8515d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 8525d4f98a2SYan Zheng 8535d4f98a2SYan Zheng node = kmalloc(sizeof(*node), GFP_NOFS); 8545d4f98a2SYan Zheng BUG_ON(!node); 8555d4f98a2SYan Zheng 8565d4f98a2SYan Zheng node->bytenr = root->node->start; 8575d4f98a2SYan Zheng node->data = root; 8585d4f98a2SYan Zheng 8595d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 8605d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 8615d4f98a2SYan Zheng node->bytenr, &node->rb_node); 8625d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 8635d4f98a2SYan Zheng BUG_ON(rb_node); 8645d4f98a2SYan Zheng 8655d4f98a2SYan Zheng list_add_tail(&root->root_list, &rc->reloc_roots); 8665d4f98a2SYan Zheng return 0; 8675d4f98a2SYan Zheng } 8685d4f98a2SYan Zheng 8695d4f98a2SYan Zheng /* 8705d4f98a2SYan Zheng * helper to update/delete the 'address of tree root -> reloc tree' 8715d4f98a2SYan Zheng * mapping 8725d4f98a2SYan Zheng */ 8735d4f98a2SYan Zheng static int __update_reloc_root(struct btrfs_root *root, int del) 8745d4f98a2SYan Zheng { 8755d4f98a2SYan Zheng struct rb_node *rb_node; 8765d4f98a2SYan Zheng struct mapping_node *node = NULL; 8775d4f98a2SYan Zheng struct reloc_control *rc = root->fs_info->reloc_ctl; 8785d4f98a2SYan Zheng 8795d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 8805d4f98a2SYan Zheng rb_node = tree_search(&rc->reloc_root_tree.rb_root, 8815d4f98a2SYan Zheng root->commit_root->start); 8825d4f98a2SYan Zheng if (rb_node) { 8835d4f98a2SYan Zheng node = rb_entry(rb_node, struct mapping_node, rb_node); 8845d4f98a2SYan Zheng rb_erase(&node->rb_node, &rc->reloc_root_tree.rb_root); 8855d4f98a2SYan Zheng } 8865d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 8875d4f98a2SYan Zheng 8885d4f98a2SYan Zheng BUG_ON((struct btrfs_root *)node->data != root); 8895d4f98a2SYan Zheng 8905d4f98a2SYan Zheng if (!del) { 8915d4f98a2SYan Zheng spin_lock(&rc->reloc_root_tree.lock); 8925d4f98a2SYan Zheng node->bytenr = root->node->start; 8935d4f98a2SYan Zheng rb_node = tree_insert(&rc->reloc_root_tree.rb_root, 8945d4f98a2SYan Zheng node->bytenr, &node->rb_node); 8955d4f98a2SYan Zheng spin_unlock(&rc->reloc_root_tree.lock); 8965d4f98a2SYan Zheng BUG_ON(rb_node); 8975d4f98a2SYan Zheng } else { 8985d4f98a2SYan Zheng list_del_init(&root->root_list); 8995d4f98a2SYan Zheng kfree(node); 9005d4f98a2SYan Zheng } 9015d4f98a2SYan Zheng return 0; 9025d4f98a2SYan Zheng } 9035d4f98a2SYan Zheng 9045d4f98a2SYan Zheng /* 9055d4f98a2SYan Zheng * create reloc tree for a given fs tree. reloc tree is just a 9065d4f98a2SYan Zheng * snapshot of the fs tree with special root objectid. 9075d4f98a2SYan Zheng */ 9085d4f98a2SYan Zheng int btrfs_init_reloc_root(struct btrfs_trans_handle *trans, 9095d4f98a2SYan Zheng struct btrfs_root *root) 9105d4f98a2SYan Zheng { 9115d4f98a2SYan Zheng struct btrfs_root *reloc_root; 9125d4f98a2SYan Zheng struct extent_buffer *eb; 9135d4f98a2SYan Zheng struct btrfs_root_item *root_item; 9145d4f98a2SYan Zheng struct btrfs_key root_key; 9155d4f98a2SYan Zheng int ret; 9165d4f98a2SYan Zheng 9175d4f98a2SYan Zheng if (root->reloc_root) { 9185d4f98a2SYan Zheng reloc_root = root->reloc_root; 9195d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 9205d4f98a2SYan Zheng return 0; 9215d4f98a2SYan Zheng } 9225d4f98a2SYan Zheng 9235d4f98a2SYan Zheng if (!root->fs_info->reloc_ctl || 9245d4f98a2SYan Zheng !root->fs_info->reloc_ctl->create_reloc_root || 9255d4f98a2SYan Zheng root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 9265d4f98a2SYan Zheng return 0; 9275d4f98a2SYan Zheng 9285d4f98a2SYan Zheng root_item = kmalloc(sizeof(*root_item), GFP_NOFS); 9295d4f98a2SYan Zheng BUG_ON(!root_item); 9305d4f98a2SYan Zheng 9315d4f98a2SYan Zheng root_key.objectid = BTRFS_TREE_RELOC_OBJECTID; 9325d4f98a2SYan Zheng root_key.type = BTRFS_ROOT_ITEM_KEY; 9335d4f98a2SYan Zheng root_key.offset = root->root_key.objectid; 9345d4f98a2SYan Zheng 9355d4f98a2SYan Zheng ret = btrfs_copy_root(trans, root, root->commit_root, &eb, 9365d4f98a2SYan Zheng BTRFS_TREE_RELOC_OBJECTID); 9375d4f98a2SYan Zheng BUG_ON(ret); 9385d4f98a2SYan Zheng 9395d4f98a2SYan Zheng btrfs_set_root_last_snapshot(&root->root_item, trans->transid - 1); 9405d4f98a2SYan Zheng memcpy(root_item, &root->root_item, sizeof(*root_item)); 9415d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 1); 9425d4f98a2SYan Zheng btrfs_set_root_bytenr(root_item, eb->start); 9435d4f98a2SYan Zheng btrfs_set_root_level(root_item, btrfs_header_level(eb)); 9445d4f98a2SYan Zheng btrfs_set_root_generation(root_item, trans->transid); 9455d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, sizeof(struct btrfs_disk_key)); 9465d4f98a2SYan Zheng root_item->drop_level = 0; 9475d4f98a2SYan Zheng 9485d4f98a2SYan Zheng btrfs_tree_unlock(eb); 9495d4f98a2SYan Zheng free_extent_buffer(eb); 9505d4f98a2SYan Zheng 9515d4f98a2SYan Zheng ret = btrfs_insert_root(trans, root->fs_info->tree_root, 9525d4f98a2SYan Zheng &root_key, root_item); 9535d4f98a2SYan Zheng BUG_ON(ret); 9545d4f98a2SYan Zheng kfree(root_item); 9555d4f98a2SYan Zheng 9565d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root, 9575d4f98a2SYan Zheng &root_key); 9585d4f98a2SYan Zheng BUG_ON(IS_ERR(reloc_root)); 9595d4f98a2SYan Zheng reloc_root->last_trans = trans->transid; 9605d4f98a2SYan Zheng 9615d4f98a2SYan Zheng __add_reloc_root(reloc_root); 9625d4f98a2SYan Zheng root->reloc_root = reloc_root; 9635d4f98a2SYan Zheng return 0; 9645d4f98a2SYan Zheng } 9655d4f98a2SYan Zheng 9665d4f98a2SYan Zheng /* 9675d4f98a2SYan Zheng * update root item of reloc tree 9685d4f98a2SYan Zheng */ 9695d4f98a2SYan Zheng int btrfs_update_reloc_root(struct btrfs_trans_handle *trans, 9705d4f98a2SYan Zheng struct btrfs_root *root) 9715d4f98a2SYan Zheng { 9725d4f98a2SYan Zheng struct btrfs_root *reloc_root; 9735d4f98a2SYan Zheng struct btrfs_root_item *root_item; 9745d4f98a2SYan Zheng int del = 0; 9755d4f98a2SYan Zheng int ret; 9765d4f98a2SYan Zheng 9775d4f98a2SYan Zheng if (!root->reloc_root) 9785d4f98a2SYan Zheng return 0; 9795d4f98a2SYan Zheng 9805d4f98a2SYan Zheng reloc_root = root->reloc_root; 9815d4f98a2SYan Zheng root_item = &reloc_root->root_item; 9825d4f98a2SYan Zheng 9835d4f98a2SYan Zheng if (btrfs_root_refs(root_item) == 0) { 9845d4f98a2SYan Zheng root->reloc_root = NULL; 9855d4f98a2SYan Zheng del = 1; 9865d4f98a2SYan Zheng } 9875d4f98a2SYan Zheng 9885d4f98a2SYan Zheng __update_reloc_root(reloc_root, del); 9895d4f98a2SYan Zheng 9905d4f98a2SYan Zheng if (reloc_root->commit_root != reloc_root->node) { 9915d4f98a2SYan Zheng btrfs_set_root_node(root_item, reloc_root->node); 9925d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 9935d4f98a2SYan Zheng reloc_root->commit_root = btrfs_root_node(reloc_root); 9945d4f98a2SYan Zheng } 9955d4f98a2SYan Zheng 9965d4f98a2SYan Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 9975d4f98a2SYan Zheng &reloc_root->root_key, root_item); 9985d4f98a2SYan Zheng BUG_ON(ret); 9995d4f98a2SYan Zheng return 0; 10005d4f98a2SYan Zheng } 10015d4f98a2SYan Zheng 10025d4f98a2SYan Zheng /* 10035d4f98a2SYan Zheng * helper to find first cached inode with inode number >= objectid 10045d4f98a2SYan Zheng * in a subvolume 10055d4f98a2SYan Zheng */ 10065d4f98a2SYan Zheng static struct inode *find_next_inode(struct btrfs_root *root, u64 objectid) 10075d4f98a2SYan Zheng { 10085d4f98a2SYan Zheng struct rb_node *node; 10095d4f98a2SYan Zheng struct rb_node *prev; 10105d4f98a2SYan Zheng struct btrfs_inode *entry; 10115d4f98a2SYan Zheng struct inode *inode; 10125d4f98a2SYan Zheng 10135d4f98a2SYan Zheng spin_lock(&root->inode_lock); 10145d4f98a2SYan Zheng again: 10155d4f98a2SYan Zheng node = root->inode_tree.rb_node; 10165d4f98a2SYan Zheng prev = NULL; 10175d4f98a2SYan Zheng while (node) { 10185d4f98a2SYan Zheng prev = node; 10195d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 10205d4f98a2SYan Zheng 10215d4f98a2SYan Zheng if (objectid < entry->vfs_inode.i_ino) 10225d4f98a2SYan Zheng node = node->rb_left; 10235d4f98a2SYan Zheng else if (objectid > entry->vfs_inode.i_ino) 10245d4f98a2SYan Zheng node = node->rb_right; 10255d4f98a2SYan Zheng else 10265d4f98a2SYan Zheng break; 10275d4f98a2SYan Zheng } 10285d4f98a2SYan Zheng if (!node) { 10295d4f98a2SYan Zheng while (prev) { 10305d4f98a2SYan Zheng entry = rb_entry(prev, struct btrfs_inode, rb_node); 10315d4f98a2SYan Zheng if (objectid <= entry->vfs_inode.i_ino) { 10325d4f98a2SYan Zheng node = prev; 10335d4f98a2SYan Zheng break; 10345d4f98a2SYan Zheng } 10355d4f98a2SYan Zheng prev = rb_next(prev); 10365d4f98a2SYan Zheng } 10375d4f98a2SYan Zheng } 10385d4f98a2SYan Zheng while (node) { 10395d4f98a2SYan Zheng entry = rb_entry(node, struct btrfs_inode, rb_node); 10405d4f98a2SYan Zheng inode = igrab(&entry->vfs_inode); 10415d4f98a2SYan Zheng if (inode) { 10425d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 10435d4f98a2SYan Zheng return inode; 10445d4f98a2SYan Zheng } 10455d4f98a2SYan Zheng 10465d4f98a2SYan Zheng objectid = entry->vfs_inode.i_ino + 1; 10475d4f98a2SYan Zheng if (cond_resched_lock(&root->inode_lock)) 10485d4f98a2SYan Zheng goto again; 10495d4f98a2SYan Zheng 10505d4f98a2SYan Zheng node = rb_next(node); 10515d4f98a2SYan Zheng } 10525d4f98a2SYan Zheng spin_unlock(&root->inode_lock); 10535d4f98a2SYan Zheng return NULL; 10545d4f98a2SYan Zheng } 10555d4f98a2SYan Zheng 10565d4f98a2SYan Zheng static int in_block_group(u64 bytenr, 10575d4f98a2SYan Zheng struct btrfs_block_group_cache *block_group) 10585d4f98a2SYan Zheng { 10595d4f98a2SYan Zheng if (bytenr >= block_group->key.objectid && 10605d4f98a2SYan Zheng bytenr < block_group->key.objectid + block_group->key.offset) 10615d4f98a2SYan Zheng return 1; 10625d4f98a2SYan Zheng return 0; 10635d4f98a2SYan Zheng } 10645d4f98a2SYan Zheng 10655d4f98a2SYan Zheng /* 10665d4f98a2SYan Zheng * get new location of data 10675d4f98a2SYan Zheng */ 10685d4f98a2SYan Zheng static int get_new_location(struct inode *reloc_inode, u64 *new_bytenr, 10695d4f98a2SYan Zheng u64 bytenr, u64 num_bytes) 10705d4f98a2SYan Zheng { 10715d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(reloc_inode)->root; 10725d4f98a2SYan Zheng struct btrfs_path *path; 10735d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 10745d4f98a2SYan Zheng struct extent_buffer *leaf; 10755d4f98a2SYan Zheng int ret; 10765d4f98a2SYan Zheng 10775d4f98a2SYan Zheng path = btrfs_alloc_path(); 10785d4f98a2SYan Zheng if (!path) 10795d4f98a2SYan Zheng return -ENOMEM; 10805d4f98a2SYan Zheng 10815d4f98a2SYan Zheng bytenr -= BTRFS_I(reloc_inode)->index_cnt; 10825d4f98a2SYan Zheng ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino, 10835d4f98a2SYan Zheng bytenr, 0); 10845d4f98a2SYan Zheng if (ret < 0) 10855d4f98a2SYan Zheng goto out; 10865d4f98a2SYan Zheng if (ret > 0) { 10875d4f98a2SYan Zheng ret = -ENOENT; 10885d4f98a2SYan Zheng goto out; 10895d4f98a2SYan Zheng } 10905d4f98a2SYan Zheng 10915d4f98a2SYan Zheng leaf = path->nodes[0]; 10925d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 10935d4f98a2SYan Zheng struct btrfs_file_extent_item); 10945d4f98a2SYan Zheng 10955d4f98a2SYan Zheng BUG_ON(btrfs_file_extent_offset(leaf, fi) || 10965d4f98a2SYan Zheng btrfs_file_extent_compression(leaf, fi) || 10975d4f98a2SYan Zheng btrfs_file_extent_encryption(leaf, fi) || 10985d4f98a2SYan Zheng btrfs_file_extent_other_encoding(leaf, fi)); 10995d4f98a2SYan Zheng 11005d4f98a2SYan Zheng if (num_bytes != btrfs_file_extent_disk_num_bytes(leaf, fi)) { 11015d4f98a2SYan Zheng ret = 1; 11025d4f98a2SYan Zheng goto out; 11035d4f98a2SYan Zheng } 11045d4f98a2SYan Zheng 11055d4f98a2SYan Zheng if (new_bytenr) 11065d4f98a2SYan Zheng *new_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 11075d4f98a2SYan Zheng ret = 0; 11085d4f98a2SYan Zheng out: 11095d4f98a2SYan Zheng btrfs_free_path(path); 11105d4f98a2SYan Zheng return ret; 11115d4f98a2SYan Zheng } 11125d4f98a2SYan Zheng 11135d4f98a2SYan Zheng /* 11145d4f98a2SYan Zheng * update file extent items in the tree leaf to point to 11155d4f98a2SYan Zheng * the new locations. 11165d4f98a2SYan Zheng */ 11175d4f98a2SYan Zheng static int replace_file_extents(struct btrfs_trans_handle *trans, 11185d4f98a2SYan Zheng struct reloc_control *rc, 11195d4f98a2SYan Zheng struct btrfs_root *root, 11205d4f98a2SYan Zheng struct extent_buffer *leaf, 11215d4f98a2SYan Zheng struct list_head *inode_list) 11225d4f98a2SYan Zheng { 11235d4f98a2SYan Zheng struct btrfs_key key; 11245d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 11255d4f98a2SYan Zheng struct inode *inode = NULL; 11265d4f98a2SYan Zheng struct inodevec *ivec = NULL; 11275d4f98a2SYan Zheng u64 parent; 11285d4f98a2SYan Zheng u64 bytenr; 11295d4f98a2SYan Zheng u64 new_bytenr; 11305d4f98a2SYan Zheng u64 num_bytes; 11315d4f98a2SYan Zheng u64 end; 11325d4f98a2SYan Zheng u32 nritems; 11335d4f98a2SYan Zheng u32 i; 11345d4f98a2SYan Zheng int ret; 11355d4f98a2SYan Zheng int first = 1; 11365d4f98a2SYan Zheng int dirty = 0; 11375d4f98a2SYan Zheng 11385d4f98a2SYan Zheng if (rc->stage != UPDATE_DATA_PTRS) 11395d4f98a2SYan Zheng return 0; 11405d4f98a2SYan Zheng 11415d4f98a2SYan Zheng /* reloc trees always use full backref */ 11425d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) 11435d4f98a2SYan Zheng parent = leaf->start; 11445d4f98a2SYan Zheng else 11455d4f98a2SYan Zheng parent = 0; 11465d4f98a2SYan Zheng 11475d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 11485d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 11495d4f98a2SYan Zheng cond_resched(); 11505d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, i); 11515d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_DATA_KEY) 11525d4f98a2SYan Zheng continue; 11535d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 11545d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 11555d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 11565d4f98a2SYan Zheng continue; 11575d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 11585d4f98a2SYan Zheng num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi); 11595d4f98a2SYan Zheng if (bytenr == 0) 11605d4f98a2SYan Zheng continue; 11615d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group)) 11625d4f98a2SYan Zheng continue; 11635d4f98a2SYan Zheng 11645d4f98a2SYan Zheng /* 11655d4f98a2SYan Zheng * if we are modifying block in fs tree, wait for readpage 11665d4f98a2SYan Zheng * to complete and drop the extent cache 11675d4f98a2SYan Zheng */ 11685d4f98a2SYan Zheng if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) { 11695d4f98a2SYan Zheng if (!ivec || ivec->nr == INODEVEC_SIZE) { 11705d4f98a2SYan Zheng ivec = kmalloc(sizeof(*ivec), GFP_NOFS); 11715d4f98a2SYan Zheng BUG_ON(!ivec); 11725d4f98a2SYan Zheng ivec->nr = 0; 11735d4f98a2SYan Zheng list_add_tail(&ivec->list, inode_list); 11745d4f98a2SYan Zheng } 11755d4f98a2SYan Zheng if (first) { 11765d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11775d4f98a2SYan Zheng if (inode) 11785d4f98a2SYan Zheng ivec->inode[ivec->nr++] = inode; 11795d4f98a2SYan Zheng first = 0; 11805d4f98a2SYan Zheng } else if (inode && inode->i_ino < key.objectid) { 11815d4f98a2SYan Zheng inode = find_next_inode(root, key.objectid); 11825d4f98a2SYan Zheng if (inode) 11835d4f98a2SYan Zheng ivec->inode[ivec->nr++] = inode; 11845d4f98a2SYan Zheng } 11855d4f98a2SYan Zheng if (inode && inode->i_ino == key.objectid) { 11865d4f98a2SYan Zheng end = key.offset + 11875d4f98a2SYan Zheng btrfs_file_extent_num_bytes(leaf, fi); 11885d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(key.offset, 11895d4f98a2SYan Zheng root->sectorsize)); 11905d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 11915d4f98a2SYan Zheng end--; 11925d4f98a2SYan Zheng ret = try_lock_extent(&BTRFS_I(inode)->io_tree, 11935d4f98a2SYan Zheng key.offset, end, 11945d4f98a2SYan Zheng GFP_NOFS); 11955d4f98a2SYan Zheng if (!ret) 11965d4f98a2SYan Zheng continue; 11975d4f98a2SYan Zheng 11985d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, key.offset, end, 11995d4f98a2SYan Zheng 1); 12005d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 12015d4f98a2SYan Zheng key.offset, end, GFP_NOFS); 12025d4f98a2SYan Zheng } 12035d4f98a2SYan Zheng } 12045d4f98a2SYan Zheng 12055d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, &new_bytenr, 12065d4f98a2SYan Zheng bytenr, num_bytes); 12075d4f98a2SYan Zheng if (ret > 0) 12085d4f98a2SYan Zheng continue; 12095d4f98a2SYan Zheng BUG_ON(ret < 0); 12105d4f98a2SYan Zheng 12115d4f98a2SYan Zheng btrfs_set_file_extent_disk_bytenr(leaf, fi, new_bytenr); 12125d4f98a2SYan Zheng dirty = 1; 12135d4f98a2SYan Zheng 12145d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 12155d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, new_bytenr, 12165d4f98a2SYan Zheng num_bytes, parent, 12175d4f98a2SYan Zheng btrfs_header_owner(leaf), 12185d4f98a2SYan Zheng key.objectid, key.offset); 12195d4f98a2SYan Zheng BUG_ON(ret); 12205d4f98a2SYan Zheng 12215d4f98a2SYan Zheng ret = btrfs_free_extent(trans, root, bytenr, num_bytes, 12225d4f98a2SYan Zheng parent, btrfs_header_owner(leaf), 12235d4f98a2SYan Zheng key.objectid, key.offset); 12245d4f98a2SYan Zheng BUG_ON(ret); 12255d4f98a2SYan Zheng } 12265d4f98a2SYan Zheng if (dirty) 12275d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 12285d4f98a2SYan Zheng return 0; 12295d4f98a2SYan Zheng } 12305d4f98a2SYan Zheng 12315d4f98a2SYan Zheng static noinline_for_stack 12325d4f98a2SYan Zheng int memcmp_node_keys(struct extent_buffer *eb, int slot, 12335d4f98a2SYan Zheng struct btrfs_path *path, int level) 12345d4f98a2SYan Zheng { 12355d4f98a2SYan Zheng struct btrfs_disk_key key1; 12365d4f98a2SYan Zheng struct btrfs_disk_key key2; 12375d4f98a2SYan Zheng btrfs_node_key(eb, &key1, slot); 12385d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &key2, path->slots[level]); 12395d4f98a2SYan Zheng return memcmp(&key1, &key2, sizeof(key1)); 12405d4f98a2SYan Zheng } 12415d4f98a2SYan Zheng 12425d4f98a2SYan Zheng /* 12435d4f98a2SYan Zheng * try to replace tree blocks in fs tree with the new blocks 12445d4f98a2SYan Zheng * in reloc tree. tree blocks haven't been modified since the 12455d4f98a2SYan Zheng * reloc tree was create can be replaced. 12465d4f98a2SYan Zheng * 12475d4f98a2SYan Zheng * if a block was replaced, level of the block + 1 is returned. 12485d4f98a2SYan Zheng * if no block got replaced, 0 is returned. if there are other 12495d4f98a2SYan Zheng * errors, a negative error number is returned. 12505d4f98a2SYan Zheng */ 12515d4f98a2SYan Zheng static int replace_path(struct btrfs_trans_handle *trans, 12525d4f98a2SYan Zheng struct btrfs_root *dest, struct btrfs_root *src, 12535d4f98a2SYan Zheng struct btrfs_path *path, struct btrfs_key *next_key, 12545d4f98a2SYan Zheng struct extent_buffer **leaf, 12555d4f98a2SYan Zheng int lowest_level, int max_level) 12565d4f98a2SYan Zheng { 12575d4f98a2SYan Zheng struct extent_buffer *eb; 12585d4f98a2SYan Zheng struct extent_buffer *parent; 12595d4f98a2SYan Zheng struct btrfs_key key; 12605d4f98a2SYan Zheng u64 old_bytenr; 12615d4f98a2SYan Zheng u64 new_bytenr; 12625d4f98a2SYan Zheng u64 old_ptr_gen; 12635d4f98a2SYan Zheng u64 new_ptr_gen; 12645d4f98a2SYan Zheng u64 last_snapshot; 12655d4f98a2SYan Zheng u32 blocksize; 12665d4f98a2SYan Zheng int level; 12675d4f98a2SYan Zheng int ret; 12685d4f98a2SYan Zheng int slot; 12695d4f98a2SYan Zheng 12705d4f98a2SYan Zheng BUG_ON(src->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID); 12715d4f98a2SYan Zheng BUG_ON(dest->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID); 12725d4f98a2SYan Zheng BUG_ON(lowest_level > 1 && leaf); 12735d4f98a2SYan Zheng 12745d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&src->root_item); 12755d4f98a2SYan Zheng 12765d4f98a2SYan Zheng slot = path->slots[lowest_level]; 12775d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[lowest_level], &key, slot); 12785d4f98a2SYan Zheng 12795d4f98a2SYan Zheng eb = btrfs_lock_root_node(dest); 12805d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 12815d4f98a2SYan Zheng level = btrfs_header_level(eb); 12825d4f98a2SYan Zheng 12835d4f98a2SYan Zheng if (level < lowest_level) { 12845d4f98a2SYan Zheng btrfs_tree_unlock(eb); 12855d4f98a2SYan Zheng free_extent_buffer(eb); 12865d4f98a2SYan Zheng return 0; 12875d4f98a2SYan Zheng } 12885d4f98a2SYan Zheng 12895d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, NULL, 0, &eb); 12905d4f98a2SYan Zheng BUG_ON(ret); 12915d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 12925d4f98a2SYan Zheng 12935d4f98a2SYan Zheng if (next_key) { 12945d4f98a2SYan Zheng next_key->objectid = (u64)-1; 12955d4f98a2SYan Zheng next_key->type = (u8)-1; 12965d4f98a2SYan Zheng next_key->offset = (u64)-1; 12975d4f98a2SYan Zheng } 12985d4f98a2SYan Zheng 12995d4f98a2SYan Zheng parent = eb; 13005d4f98a2SYan Zheng while (1) { 13015d4f98a2SYan Zheng level = btrfs_header_level(parent); 13025d4f98a2SYan Zheng BUG_ON(level < lowest_level); 13035d4f98a2SYan Zheng 13045d4f98a2SYan Zheng ret = btrfs_bin_search(parent, &key, level, &slot); 13055d4f98a2SYan Zheng if (ret && slot > 0) 13065d4f98a2SYan Zheng slot--; 13075d4f98a2SYan Zheng 13085d4f98a2SYan Zheng if (next_key && slot + 1 < btrfs_header_nritems(parent)) 13095d4f98a2SYan Zheng btrfs_node_key_to_cpu(parent, next_key, slot + 1); 13105d4f98a2SYan Zheng 13115d4f98a2SYan Zheng old_bytenr = btrfs_node_blockptr(parent, slot); 13125d4f98a2SYan Zheng blocksize = btrfs_level_size(dest, level - 1); 13135d4f98a2SYan Zheng old_ptr_gen = btrfs_node_ptr_generation(parent, slot); 13145d4f98a2SYan Zheng 13155d4f98a2SYan Zheng if (level <= max_level) { 13165d4f98a2SYan Zheng eb = path->nodes[level]; 13175d4f98a2SYan Zheng new_bytenr = btrfs_node_blockptr(eb, 13185d4f98a2SYan Zheng path->slots[level]); 13195d4f98a2SYan Zheng new_ptr_gen = btrfs_node_ptr_generation(eb, 13205d4f98a2SYan Zheng path->slots[level]); 13215d4f98a2SYan Zheng } else { 13225d4f98a2SYan Zheng new_bytenr = 0; 13235d4f98a2SYan Zheng new_ptr_gen = 0; 13245d4f98a2SYan Zheng } 13255d4f98a2SYan Zheng 13265d4f98a2SYan Zheng if (new_bytenr > 0 && new_bytenr == old_bytenr) { 13275d4f98a2SYan Zheng WARN_ON(1); 13285d4f98a2SYan Zheng ret = level; 13295d4f98a2SYan Zheng break; 13305d4f98a2SYan Zheng } 13315d4f98a2SYan Zheng 13325d4f98a2SYan Zheng if (new_bytenr == 0 || old_ptr_gen > last_snapshot || 13335d4f98a2SYan Zheng memcmp_node_keys(parent, slot, path, level)) { 13345d4f98a2SYan Zheng if (level <= lowest_level && !leaf) { 13355d4f98a2SYan Zheng ret = 0; 13365d4f98a2SYan Zheng break; 13375d4f98a2SYan Zheng } 13385d4f98a2SYan Zheng 13395d4f98a2SYan Zheng eb = read_tree_block(dest, old_bytenr, blocksize, 13405d4f98a2SYan Zheng old_ptr_gen); 13415d4f98a2SYan Zheng btrfs_tree_lock(eb); 13425d4f98a2SYan Zheng ret = btrfs_cow_block(trans, dest, eb, parent, 13435d4f98a2SYan Zheng slot, &eb); 13445d4f98a2SYan Zheng BUG_ON(ret); 13455d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 13465d4f98a2SYan Zheng 13475d4f98a2SYan Zheng if (level <= lowest_level) { 13485d4f98a2SYan Zheng *leaf = eb; 13495d4f98a2SYan Zheng ret = 0; 13505d4f98a2SYan Zheng break; 13515d4f98a2SYan Zheng } 13525d4f98a2SYan Zheng 13535d4f98a2SYan Zheng btrfs_tree_unlock(parent); 13545d4f98a2SYan Zheng free_extent_buffer(parent); 13555d4f98a2SYan Zheng 13565d4f98a2SYan Zheng parent = eb; 13575d4f98a2SYan Zheng continue; 13585d4f98a2SYan Zheng } 13595d4f98a2SYan Zheng 13605d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 13615d4f98a2SYan Zheng path->slots[level]); 13625d4f98a2SYan Zheng btrfs_release_path(src, path); 13635d4f98a2SYan Zheng 13645d4f98a2SYan Zheng path->lowest_level = level; 13655d4f98a2SYan Zheng ret = btrfs_search_slot(trans, src, &key, path, 0, 1); 13665d4f98a2SYan Zheng path->lowest_level = 0; 13675d4f98a2SYan Zheng BUG_ON(ret); 13685d4f98a2SYan Zheng 13695d4f98a2SYan Zheng /* 13705d4f98a2SYan Zheng * swap blocks in fs tree and reloc tree. 13715d4f98a2SYan Zheng */ 13725d4f98a2SYan Zheng btrfs_set_node_blockptr(parent, slot, new_bytenr); 13735d4f98a2SYan Zheng btrfs_set_node_ptr_generation(parent, slot, new_ptr_gen); 13745d4f98a2SYan Zheng btrfs_mark_buffer_dirty(parent); 13755d4f98a2SYan Zheng 13765d4f98a2SYan Zheng btrfs_set_node_blockptr(path->nodes[level], 13775d4f98a2SYan Zheng path->slots[level], old_bytenr); 13785d4f98a2SYan Zheng btrfs_set_node_ptr_generation(path->nodes[level], 13795d4f98a2SYan Zheng path->slots[level], old_ptr_gen); 13805d4f98a2SYan Zheng btrfs_mark_buffer_dirty(path->nodes[level]); 13815d4f98a2SYan Zheng 13825d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, src, old_bytenr, blocksize, 13835d4f98a2SYan Zheng path->nodes[level]->start, 13845d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 13855d4f98a2SYan Zheng BUG_ON(ret); 13865d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, dest, new_bytenr, blocksize, 13875d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 13885d4f98a2SYan Zheng 0); 13895d4f98a2SYan Zheng BUG_ON(ret); 13905d4f98a2SYan Zheng 13915d4f98a2SYan Zheng ret = btrfs_free_extent(trans, src, new_bytenr, blocksize, 13925d4f98a2SYan Zheng path->nodes[level]->start, 13935d4f98a2SYan Zheng src->root_key.objectid, level - 1, 0); 13945d4f98a2SYan Zheng BUG_ON(ret); 13955d4f98a2SYan Zheng 13965d4f98a2SYan Zheng ret = btrfs_free_extent(trans, dest, old_bytenr, blocksize, 13975d4f98a2SYan Zheng 0, dest->root_key.objectid, level - 1, 13985d4f98a2SYan Zheng 0); 13995d4f98a2SYan Zheng BUG_ON(ret); 14005d4f98a2SYan Zheng 14015d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 14025d4f98a2SYan Zheng 14035d4f98a2SYan Zheng ret = level; 14045d4f98a2SYan Zheng break; 14055d4f98a2SYan Zheng } 14065d4f98a2SYan Zheng btrfs_tree_unlock(parent); 14075d4f98a2SYan Zheng free_extent_buffer(parent); 14085d4f98a2SYan Zheng return ret; 14095d4f98a2SYan Zheng } 14105d4f98a2SYan Zheng 14115d4f98a2SYan Zheng /* 14125d4f98a2SYan Zheng * helper to find next relocated block in reloc tree 14135d4f98a2SYan Zheng */ 14145d4f98a2SYan Zheng static noinline_for_stack 14155d4f98a2SYan Zheng int walk_up_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14165d4f98a2SYan Zheng int *level) 14175d4f98a2SYan Zheng { 14185d4f98a2SYan Zheng struct extent_buffer *eb; 14195d4f98a2SYan Zheng int i; 14205d4f98a2SYan Zheng u64 last_snapshot; 14215d4f98a2SYan Zheng u32 nritems; 14225d4f98a2SYan Zheng 14235d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14245d4f98a2SYan Zheng 14255d4f98a2SYan Zheng for (i = 0; i < *level; i++) { 14265d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14275d4f98a2SYan Zheng path->nodes[i] = NULL; 14285d4f98a2SYan Zheng } 14295d4f98a2SYan Zheng 14305d4f98a2SYan Zheng for (i = *level; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) { 14315d4f98a2SYan Zheng eb = path->nodes[i]; 14325d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14335d4f98a2SYan Zheng while (path->slots[i] + 1 < nritems) { 14345d4f98a2SYan Zheng path->slots[i]++; 14355d4f98a2SYan Zheng if (btrfs_node_ptr_generation(eb, path->slots[i]) <= 14365d4f98a2SYan Zheng last_snapshot) 14375d4f98a2SYan Zheng continue; 14385d4f98a2SYan Zheng 14395d4f98a2SYan Zheng *level = i; 14405d4f98a2SYan Zheng return 0; 14415d4f98a2SYan Zheng } 14425d4f98a2SYan Zheng free_extent_buffer(path->nodes[i]); 14435d4f98a2SYan Zheng path->nodes[i] = NULL; 14445d4f98a2SYan Zheng } 14455d4f98a2SYan Zheng return 1; 14465d4f98a2SYan Zheng } 14475d4f98a2SYan Zheng 14485d4f98a2SYan Zheng /* 14495d4f98a2SYan Zheng * walk down reloc tree to find relocated block of lowest level 14505d4f98a2SYan Zheng */ 14515d4f98a2SYan Zheng static noinline_for_stack 14525d4f98a2SYan Zheng int walk_down_reloc_tree(struct btrfs_root *root, struct btrfs_path *path, 14535d4f98a2SYan Zheng int *level) 14545d4f98a2SYan Zheng { 14555d4f98a2SYan Zheng struct extent_buffer *eb = NULL; 14565d4f98a2SYan Zheng int i; 14575d4f98a2SYan Zheng u64 bytenr; 14585d4f98a2SYan Zheng u64 ptr_gen = 0; 14595d4f98a2SYan Zheng u64 last_snapshot; 14605d4f98a2SYan Zheng u32 blocksize; 14615d4f98a2SYan Zheng u32 nritems; 14625d4f98a2SYan Zheng 14635d4f98a2SYan Zheng last_snapshot = btrfs_root_last_snapshot(&root->root_item); 14645d4f98a2SYan Zheng 14655d4f98a2SYan Zheng for (i = *level; i > 0; i--) { 14665d4f98a2SYan Zheng eb = path->nodes[i]; 14675d4f98a2SYan Zheng nritems = btrfs_header_nritems(eb); 14685d4f98a2SYan Zheng while (path->slots[i] < nritems) { 14695d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(eb, path->slots[i]); 14705d4f98a2SYan Zheng if (ptr_gen > last_snapshot) 14715d4f98a2SYan Zheng break; 14725d4f98a2SYan Zheng path->slots[i]++; 14735d4f98a2SYan Zheng } 14745d4f98a2SYan Zheng if (path->slots[i] >= nritems) { 14755d4f98a2SYan Zheng if (i == *level) 14765d4f98a2SYan Zheng break; 14775d4f98a2SYan Zheng *level = i + 1; 14785d4f98a2SYan Zheng return 0; 14795d4f98a2SYan Zheng } 14805d4f98a2SYan Zheng if (i == 1) { 14815d4f98a2SYan Zheng *level = i; 14825d4f98a2SYan Zheng return 0; 14835d4f98a2SYan Zheng } 14845d4f98a2SYan Zheng 14855d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(eb, path->slots[i]); 14865d4f98a2SYan Zheng blocksize = btrfs_level_size(root, i - 1); 14875d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, ptr_gen); 14885d4f98a2SYan Zheng BUG_ON(btrfs_header_level(eb) != i - 1); 14895d4f98a2SYan Zheng path->nodes[i - 1] = eb; 14905d4f98a2SYan Zheng path->slots[i - 1] = 0; 14915d4f98a2SYan Zheng } 14925d4f98a2SYan Zheng return 1; 14935d4f98a2SYan Zheng } 14945d4f98a2SYan Zheng 14955d4f98a2SYan Zheng /* 14965d4f98a2SYan Zheng * invalidate extent cache for file extents whose key in range of 14975d4f98a2SYan Zheng * [min_key, max_key) 14985d4f98a2SYan Zheng */ 14995d4f98a2SYan Zheng static int invalidate_extent_cache(struct btrfs_root *root, 15005d4f98a2SYan Zheng struct btrfs_key *min_key, 15015d4f98a2SYan Zheng struct btrfs_key *max_key) 15025d4f98a2SYan Zheng { 15035d4f98a2SYan Zheng struct inode *inode = NULL; 15045d4f98a2SYan Zheng u64 objectid; 15055d4f98a2SYan Zheng u64 start, end; 15065d4f98a2SYan Zheng 15075d4f98a2SYan Zheng objectid = min_key->objectid; 15085d4f98a2SYan Zheng while (1) { 15095d4f98a2SYan Zheng cond_resched(); 15105d4f98a2SYan Zheng iput(inode); 15115d4f98a2SYan Zheng 15125d4f98a2SYan Zheng if (objectid > max_key->objectid) 15135d4f98a2SYan Zheng break; 15145d4f98a2SYan Zheng 15155d4f98a2SYan Zheng inode = find_next_inode(root, objectid); 15165d4f98a2SYan Zheng if (!inode) 15175d4f98a2SYan Zheng break; 15185d4f98a2SYan Zheng 15195d4f98a2SYan Zheng if (inode->i_ino > max_key->objectid) { 15205d4f98a2SYan Zheng iput(inode); 15215d4f98a2SYan Zheng break; 15225d4f98a2SYan Zheng } 15235d4f98a2SYan Zheng 15245d4f98a2SYan Zheng objectid = inode->i_ino + 1; 15255d4f98a2SYan Zheng if (!S_ISREG(inode->i_mode)) 15265d4f98a2SYan Zheng continue; 15275d4f98a2SYan Zheng 15285d4f98a2SYan Zheng if (unlikely(min_key->objectid == inode->i_ino)) { 15295d4f98a2SYan Zheng if (min_key->type > BTRFS_EXTENT_DATA_KEY) 15305d4f98a2SYan Zheng continue; 15315d4f98a2SYan Zheng if (min_key->type < BTRFS_EXTENT_DATA_KEY) 15325d4f98a2SYan Zheng start = 0; 15335d4f98a2SYan Zheng else { 15345d4f98a2SYan Zheng start = min_key->offset; 15355d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(start, root->sectorsize)); 15365d4f98a2SYan Zheng } 15375d4f98a2SYan Zheng } else { 15385d4f98a2SYan Zheng start = 0; 15395d4f98a2SYan Zheng } 15405d4f98a2SYan Zheng 15415d4f98a2SYan Zheng if (unlikely(max_key->objectid == inode->i_ino)) { 15425d4f98a2SYan Zheng if (max_key->type < BTRFS_EXTENT_DATA_KEY) 15435d4f98a2SYan Zheng continue; 15445d4f98a2SYan Zheng if (max_key->type > BTRFS_EXTENT_DATA_KEY) { 15455d4f98a2SYan Zheng end = (u64)-1; 15465d4f98a2SYan Zheng } else { 15475d4f98a2SYan Zheng if (max_key->offset == 0) 15485d4f98a2SYan Zheng continue; 15495d4f98a2SYan Zheng end = max_key->offset; 15505d4f98a2SYan Zheng WARN_ON(!IS_ALIGNED(end, root->sectorsize)); 15515d4f98a2SYan Zheng end--; 15525d4f98a2SYan Zheng } 15535d4f98a2SYan Zheng } else { 15545d4f98a2SYan Zheng end = (u64)-1; 15555d4f98a2SYan Zheng } 15565d4f98a2SYan Zheng 15575d4f98a2SYan Zheng /* the lock_extent waits for readpage to complete */ 15585d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 15595d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 1); 15605d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 15615d4f98a2SYan Zheng } 15625d4f98a2SYan Zheng return 0; 15635d4f98a2SYan Zheng } 15645d4f98a2SYan Zheng 15658082510eSYan, Zheng static void put_inodes(struct list_head *list) 15668082510eSYan, Zheng { 15678082510eSYan, Zheng struct inodevec *ivec; 15688082510eSYan, Zheng while (!list_empty(list)) { 15698082510eSYan, Zheng ivec = list_entry(list->next, struct inodevec, list); 15708082510eSYan, Zheng list_del(&ivec->list); 15718082510eSYan, Zheng while (ivec->nr > 0) { 15728082510eSYan, Zheng ivec->nr--; 15738082510eSYan, Zheng iput(ivec->inode[ivec->nr]); 15748082510eSYan, Zheng } 15758082510eSYan, Zheng kfree(ivec); 15768082510eSYan, Zheng } 15778082510eSYan, Zheng } 15788082510eSYan, Zheng 15795d4f98a2SYan Zheng static int find_next_key(struct btrfs_path *path, int level, 15805d4f98a2SYan Zheng struct btrfs_key *key) 15815d4f98a2SYan Zheng 15825d4f98a2SYan Zheng { 15835d4f98a2SYan Zheng while (level < BTRFS_MAX_LEVEL) { 15845d4f98a2SYan Zheng if (!path->nodes[level]) 15855d4f98a2SYan Zheng break; 15865d4f98a2SYan Zheng if (path->slots[level] + 1 < 15875d4f98a2SYan Zheng btrfs_header_nritems(path->nodes[level])) { 15885d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], key, 15895d4f98a2SYan Zheng path->slots[level] + 1); 15905d4f98a2SYan Zheng return 0; 15915d4f98a2SYan Zheng } 15925d4f98a2SYan Zheng level++; 15935d4f98a2SYan Zheng } 15945d4f98a2SYan Zheng return 1; 15955d4f98a2SYan Zheng } 15965d4f98a2SYan Zheng 15975d4f98a2SYan Zheng /* 15985d4f98a2SYan Zheng * merge the relocated tree blocks in reloc tree with corresponding 15995d4f98a2SYan Zheng * fs tree. 16005d4f98a2SYan Zheng */ 16015d4f98a2SYan Zheng static noinline_for_stack int merge_reloc_root(struct reloc_control *rc, 16025d4f98a2SYan Zheng struct btrfs_root *root) 16035d4f98a2SYan Zheng { 16045d4f98a2SYan Zheng LIST_HEAD(inode_list); 16055d4f98a2SYan Zheng struct btrfs_key key; 16065d4f98a2SYan Zheng struct btrfs_key next_key; 16075d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 16085d4f98a2SYan Zheng struct btrfs_root *reloc_root; 16095d4f98a2SYan Zheng struct btrfs_root_item *root_item; 16105d4f98a2SYan Zheng struct btrfs_path *path; 16115d4f98a2SYan Zheng struct extent_buffer *leaf = NULL; 16125d4f98a2SYan Zheng unsigned long nr; 16135d4f98a2SYan Zheng int level; 16145d4f98a2SYan Zheng int max_level; 16155d4f98a2SYan Zheng int replaced = 0; 16165d4f98a2SYan Zheng int ret; 16175d4f98a2SYan Zheng int err = 0; 16185d4f98a2SYan Zheng 16195d4f98a2SYan Zheng path = btrfs_alloc_path(); 16205d4f98a2SYan Zheng if (!path) 16215d4f98a2SYan Zheng return -ENOMEM; 16225d4f98a2SYan Zheng 16235d4f98a2SYan Zheng reloc_root = root->reloc_root; 16245d4f98a2SYan Zheng root_item = &reloc_root->root_item; 16255d4f98a2SYan Zheng 16265d4f98a2SYan Zheng if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) { 16275d4f98a2SYan Zheng level = btrfs_root_level(root_item); 16285d4f98a2SYan Zheng extent_buffer_get(reloc_root->node); 16295d4f98a2SYan Zheng path->nodes[level] = reloc_root->node; 16305d4f98a2SYan Zheng path->slots[level] = 0; 16315d4f98a2SYan Zheng } else { 16325d4f98a2SYan Zheng btrfs_disk_key_to_cpu(&key, &root_item->drop_progress); 16335d4f98a2SYan Zheng 16345d4f98a2SYan Zheng level = root_item->drop_level; 16355d4f98a2SYan Zheng BUG_ON(level == 0); 16365d4f98a2SYan Zheng path->lowest_level = level; 16375d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, reloc_root, &key, path, 0, 0); 163833c66f43SYan Zheng path->lowest_level = 0; 16395d4f98a2SYan Zheng if (ret < 0) { 16405d4f98a2SYan Zheng btrfs_free_path(path); 16415d4f98a2SYan Zheng return ret; 16425d4f98a2SYan Zheng } 16435d4f98a2SYan Zheng 16445d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &next_key, 16455d4f98a2SYan Zheng path->slots[level]); 16465d4f98a2SYan Zheng WARN_ON(memcmp(&key, &next_key, sizeof(key))); 16475d4f98a2SYan Zheng 16485d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 0); 16495d4f98a2SYan Zheng } 16505d4f98a2SYan Zheng 16515d4f98a2SYan Zheng if (level == 0 && rc->stage == UPDATE_DATA_PTRS) { 16525d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 16535d4f98a2SYan Zheng 16545d4f98a2SYan Zheng leaf = path->nodes[0]; 16555d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, 0); 16565d4f98a2SYan Zheng btrfs_release_path(reloc_root, path); 16575d4f98a2SYan Zheng 16585d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, &key, path, 0, 1); 16595d4f98a2SYan Zheng if (ret < 0) { 16605d4f98a2SYan Zheng err = ret; 16615d4f98a2SYan Zheng goto out; 16625d4f98a2SYan Zheng } 16635d4f98a2SYan Zheng 16645d4f98a2SYan Zheng leaf = path->nodes[0]; 16655d4f98a2SYan Zheng btrfs_unlock_up_safe(path, 1); 16665d4f98a2SYan Zheng ret = replace_file_extents(trans, rc, root, leaf, 16675d4f98a2SYan Zheng &inode_list); 16685d4f98a2SYan Zheng if (ret < 0) 16695d4f98a2SYan Zheng err = ret; 16705d4f98a2SYan Zheng goto out; 16715d4f98a2SYan Zheng } 16725d4f98a2SYan Zheng 16735d4f98a2SYan Zheng memset(&next_key, 0, sizeof(next_key)); 16745d4f98a2SYan Zheng 16755d4f98a2SYan Zheng while (1) { 16765d4f98a2SYan Zheng leaf = NULL; 16775d4f98a2SYan Zheng replaced = 0; 16785d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 16795d4f98a2SYan Zheng max_level = level; 16805d4f98a2SYan Zheng 16815d4f98a2SYan Zheng ret = walk_down_reloc_tree(reloc_root, path, &level); 16825d4f98a2SYan Zheng if (ret < 0) { 16835d4f98a2SYan Zheng err = ret; 16845d4f98a2SYan Zheng goto out; 16855d4f98a2SYan Zheng } 16865d4f98a2SYan Zheng if (ret > 0) 16875d4f98a2SYan Zheng break; 16885d4f98a2SYan Zheng 16895d4f98a2SYan Zheng if (!find_next_key(path, level, &key) && 16905d4f98a2SYan Zheng btrfs_comp_cpu_keys(&next_key, &key) >= 0) { 16915d4f98a2SYan Zheng ret = 0; 16925d4f98a2SYan Zheng } else if (level == 1 && rc->stage == UPDATE_DATA_PTRS) { 16935d4f98a2SYan Zheng ret = replace_path(trans, root, reloc_root, 16945d4f98a2SYan Zheng path, &next_key, &leaf, 16955d4f98a2SYan Zheng level, max_level); 16965d4f98a2SYan Zheng } else { 16975d4f98a2SYan Zheng ret = replace_path(trans, root, reloc_root, 16985d4f98a2SYan Zheng path, &next_key, NULL, 16995d4f98a2SYan Zheng level, max_level); 17005d4f98a2SYan Zheng } 17015d4f98a2SYan Zheng if (ret < 0) { 17025d4f98a2SYan Zheng err = ret; 17035d4f98a2SYan Zheng goto out; 17045d4f98a2SYan Zheng } 17055d4f98a2SYan Zheng 17065d4f98a2SYan Zheng if (ret > 0) { 17075d4f98a2SYan Zheng level = ret; 17085d4f98a2SYan Zheng btrfs_node_key_to_cpu(path->nodes[level], &key, 17095d4f98a2SYan Zheng path->slots[level]); 17105d4f98a2SYan Zheng replaced = 1; 17115d4f98a2SYan Zheng } else if (leaf) { 17125d4f98a2SYan Zheng /* 17135d4f98a2SYan Zheng * no block got replaced, try replacing file extents 17145d4f98a2SYan Zheng */ 17155d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, 0); 17165d4f98a2SYan Zheng ret = replace_file_extents(trans, rc, root, leaf, 17175d4f98a2SYan Zheng &inode_list); 17185d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 17195d4f98a2SYan Zheng free_extent_buffer(leaf); 17205d4f98a2SYan Zheng BUG_ON(ret < 0); 17215d4f98a2SYan Zheng } 17225d4f98a2SYan Zheng 17235d4f98a2SYan Zheng ret = walk_up_reloc_tree(reloc_root, path, &level); 17245d4f98a2SYan Zheng if (ret > 0) 17255d4f98a2SYan Zheng break; 17265d4f98a2SYan Zheng 17275d4f98a2SYan Zheng BUG_ON(level == 0); 17285d4f98a2SYan Zheng /* 17295d4f98a2SYan Zheng * save the merging progress in the drop_progress. 17305d4f98a2SYan Zheng * this is OK since root refs == 1 in this case. 17315d4f98a2SYan Zheng */ 17325d4f98a2SYan Zheng btrfs_node_key(path->nodes[level], &root_item->drop_progress, 17335d4f98a2SYan Zheng path->slots[level]); 17345d4f98a2SYan Zheng root_item->drop_level = level; 17355d4f98a2SYan Zheng 17365d4f98a2SYan Zheng nr = trans->blocks_used; 17375d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 17385d4f98a2SYan Zheng 17395d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 17405d4f98a2SYan Zheng 17418082510eSYan, Zheng /* 17428082510eSYan, Zheng * put inodes outside transaction, otherwise we may deadlock. 17438082510eSYan, Zheng */ 17448082510eSYan, Zheng put_inodes(&inode_list); 17458082510eSYan, Zheng 17465d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17475d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17485d4f98a2SYan Zheng } 17495d4f98a2SYan Zheng 17505d4f98a2SYan Zheng /* 17515d4f98a2SYan Zheng * handle the case only one block in the fs tree need to be 17525d4f98a2SYan Zheng * relocated and the block is tree root. 17535d4f98a2SYan Zheng */ 17545d4f98a2SYan Zheng leaf = btrfs_lock_root_node(root); 17555d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, leaf, NULL, 0, &leaf); 17565d4f98a2SYan Zheng btrfs_tree_unlock(leaf); 17575d4f98a2SYan Zheng free_extent_buffer(leaf); 17585d4f98a2SYan Zheng if (ret < 0) 17595d4f98a2SYan Zheng err = ret; 17605d4f98a2SYan Zheng out: 17615d4f98a2SYan Zheng btrfs_free_path(path); 17625d4f98a2SYan Zheng 17635d4f98a2SYan Zheng if (err == 0) { 17645d4f98a2SYan Zheng memset(&root_item->drop_progress, 0, 17655d4f98a2SYan Zheng sizeof(root_item->drop_progress)); 17665d4f98a2SYan Zheng root_item->drop_level = 0; 17675d4f98a2SYan Zheng btrfs_set_root_refs(root_item, 0); 17685d4f98a2SYan Zheng } 17695d4f98a2SYan Zheng 17705d4f98a2SYan Zheng nr = trans->blocks_used; 17715d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 17725d4f98a2SYan Zheng 17735d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 17745d4f98a2SYan Zheng 17758082510eSYan, Zheng put_inodes(&inode_list); 17765d4f98a2SYan Zheng 17775d4f98a2SYan Zheng if (replaced && rc->stage == UPDATE_DATA_PTRS) 17785d4f98a2SYan Zheng invalidate_extent_cache(root, &key, &next_key); 17795d4f98a2SYan Zheng 17805d4f98a2SYan Zheng return err; 17815d4f98a2SYan Zheng } 17825d4f98a2SYan Zheng 17835d4f98a2SYan Zheng /* 17845d4f98a2SYan Zheng * callback for the work threads. 17855d4f98a2SYan Zheng * this function merges reloc tree with corresponding fs tree, 17865d4f98a2SYan Zheng * and then drops the reloc tree. 17875d4f98a2SYan Zheng */ 17885d4f98a2SYan Zheng static void merge_func(struct btrfs_work *work) 17895d4f98a2SYan Zheng { 17905d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 17915d4f98a2SYan Zheng struct btrfs_root *root; 17925d4f98a2SYan Zheng struct btrfs_root *reloc_root; 17935d4f98a2SYan Zheng struct async_merge *async; 17945d4f98a2SYan Zheng 17955d4f98a2SYan Zheng async = container_of(work, struct async_merge, work); 17965d4f98a2SYan Zheng reloc_root = async->root; 17975d4f98a2SYan Zheng 17985d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 17995d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, 18005d4f98a2SYan Zheng reloc_root->root_key.offset); 18015d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 18025d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 18035d4f98a2SYan Zheng 18045d4f98a2SYan Zheng merge_reloc_root(async->rc, root); 18055d4f98a2SYan Zheng 18065d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 18075d4f98a2SYan Zheng btrfs_update_reloc_root(trans, root); 18085d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 18095d4f98a2SYan Zheng } 18105d4f98a2SYan Zheng 18112c47e605SYan Zheng btrfs_drop_snapshot(reloc_root, 0); 18125d4f98a2SYan Zheng 18135d4f98a2SYan Zheng if (atomic_dec_and_test(async->num_pending)) 18145d4f98a2SYan Zheng complete(async->done); 18155d4f98a2SYan Zheng 18165d4f98a2SYan Zheng kfree(async); 18175d4f98a2SYan Zheng } 18185d4f98a2SYan Zheng 18195d4f98a2SYan Zheng static int merge_reloc_roots(struct reloc_control *rc) 18205d4f98a2SYan Zheng { 18215d4f98a2SYan Zheng struct async_merge *async; 18225d4f98a2SYan Zheng struct btrfs_root *root; 18235d4f98a2SYan Zheng struct completion done; 18245d4f98a2SYan Zheng atomic_t num_pending; 18255d4f98a2SYan Zheng 18265d4f98a2SYan Zheng init_completion(&done); 18275d4f98a2SYan Zheng atomic_set(&num_pending, 1); 18285d4f98a2SYan Zheng 18295d4f98a2SYan Zheng while (!list_empty(&rc->reloc_roots)) { 18305d4f98a2SYan Zheng root = list_entry(rc->reloc_roots.next, 18315d4f98a2SYan Zheng struct btrfs_root, root_list); 18325d4f98a2SYan Zheng list_del_init(&root->root_list); 18335d4f98a2SYan Zheng 18345d4f98a2SYan Zheng async = kmalloc(sizeof(*async), GFP_NOFS); 18355d4f98a2SYan Zheng BUG_ON(!async); 18365d4f98a2SYan Zheng async->work.func = merge_func; 18375d4f98a2SYan Zheng async->work.flags = 0; 18385d4f98a2SYan Zheng async->rc = rc; 18395d4f98a2SYan Zheng async->root = root; 18405d4f98a2SYan Zheng async->done = &done; 18415d4f98a2SYan Zheng async->num_pending = &num_pending; 18425d4f98a2SYan Zheng atomic_inc(&num_pending); 18435d4f98a2SYan Zheng btrfs_queue_worker(&rc->workers, &async->work); 18445d4f98a2SYan Zheng } 18455d4f98a2SYan Zheng 18465d4f98a2SYan Zheng if (!atomic_dec_and_test(&num_pending)) 18475d4f98a2SYan Zheng wait_for_completion(&done); 18485d4f98a2SYan Zheng 18495d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&rc->reloc_root_tree.rb_root)); 18505d4f98a2SYan Zheng return 0; 18515d4f98a2SYan Zheng } 18525d4f98a2SYan Zheng 18535d4f98a2SYan Zheng static void free_block_list(struct rb_root *blocks) 18545d4f98a2SYan Zheng { 18555d4f98a2SYan Zheng struct tree_block *block; 18565d4f98a2SYan Zheng struct rb_node *rb_node; 18575d4f98a2SYan Zheng while ((rb_node = rb_first(blocks))) { 18585d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 18595d4f98a2SYan Zheng rb_erase(rb_node, blocks); 18605d4f98a2SYan Zheng kfree(block); 18615d4f98a2SYan Zheng } 18625d4f98a2SYan Zheng } 18635d4f98a2SYan Zheng 18645d4f98a2SYan Zheng static int record_reloc_root_in_trans(struct btrfs_trans_handle *trans, 18655d4f98a2SYan Zheng struct btrfs_root *reloc_root) 18665d4f98a2SYan Zheng { 18675d4f98a2SYan Zheng struct btrfs_root *root; 18685d4f98a2SYan Zheng 18695d4f98a2SYan Zheng if (reloc_root->last_trans == trans->transid) 18705d4f98a2SYan Zheng return 0; 18715d4f98a2SYan Zheng 18725d4f98a2SYan Zheng root = read_fs_root(reloc_root->fs_info, reloc_root->root_key.offset); 18735d4f98a2SYan Zheng BUG_ON(IS_ERR(root)); 18745d4f98a2SYan Zheng BUG_ON(root->reloc_root != reloc_root); 18755d4f98a2SYan Zheng 18765d4f98a2SYan Zheng return btrfs_record_root_in_trans(trans, root); 18775d4f98a2SYan Zheng } 18785d4f98a2SYan Zheng 18795d4f98a2SYan Zheng /* 18805d4f98a2SYan Zheng * select one tree from trees that references the block. 18815d4f98a2SYan Zheng * for blocks in refernce counted trees, we preper reloc tree. 18825d4f98a2SYan Zheng * if no reloc tree found and reloc_only is true, NULL is returned. 18835d4f98a2SYan Zheng */ 18845d4f98a2SYan Zheng static struct btrfs_root *__select_one_root(struct btrfs_trans_handle *trans, 18855d4f98a2SYan Zheng struct backref_node *node, 18865d4f98a2SYan Zheng struct backref_edge *edges[], 18875d4f98a2SYan Zheng int *nr, int reloc_only) 18885d4f98a2SYan Zheng { 18895d4f98a2SYan Zheng struct backref_node *next; 18905d4f98a2SYan Zheng struct btrfs_root *root; 18915d4f98a2SYan Zheng int index; 18925d4f98a2SYan Zheng int loop = 0; 18935d4f98a2SYan Zheng again: 18945d4f98a2SYan Zheng index = 0; 18955d4f98a2SYan Zheng next = node; 18965d4f98a2SYan Zheng while (1) { 18975d4f98a2SYan Zheng cond_resched(); 18985d4f98a2SYan Zheng next = walk_up_backref(next, edges, &index); 18995d4f98a2SYan Zheng root = next->root; 19005d4f98a2SYan Zheng if (!root) { 19015d4f98a2SYan Zheng BUG_ON(!node->old_root); 19025d4f98a2SYan Zheng goto skip; 19035d4f98a2SYan Zheng } 19045d4f98a2SYan Zheng 19055d4f98a2SYan Zheng /* no other choice for non-refernce counted tree */ 19065d4f98a2SYan Zheng if (!root->ref_cows) { 19075d4f98a2SYan Zheng BUG_ON(reloc_only); 19085d4f98a2SYan Zheng break; 19095d4f98a2SYan Zheng } 19105d4f98a2SYan Zheng 19115d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 19125d4f98a2SYan Zheng record_reloc_root_in_trans(trans, root); 19135d4f98a2SYan Zheng break; 19145d4f98a2SYan Zheng } 19155d4f98a2SYan Zheng 19165d4f98a2SYan Zheng if (loop) { 19175d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 19185d4f98a2SYan Zheng break; 19195d4f98a2SYan Zheng } 19205d4f98a2SYan Zheng 19215d4f98a2SYan Zheng if (reloc_only || next != node) { 19225d4f98a2SYan Zheng if (!root->reloc_root) 19235d4f98a2SYan Zheng btrfs_record_root_in_trans(trans, root); 19245d4f98a2SYan Zheng root = root->reloc_root; 19255d4f98a2SYan Zheng /* 19265d4f98a2SYan Zheng * if the reloc tree was created in current 19275d4f98a2SYan Zheng * transation, there is no node in backref tree 19285d4f98a2SYan Zheng * corresponds to the root of the reloc tree. 19295d4f98a2SYan Zheng */ 19305d4f98a2SYan Zheng if (btrfs_root_last_snapshot(&root->root_item) == 19315d4f98a2SYan Zheng trans->transid - 1) 19325d4f98a2SYan Zheng break; 19335d4f98a2SYan Zheng } 19345d4f98a2SYan Zheng skip: 19355d4f98a2SYan Zheng root = NULL; 19365d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 19375d4f98a2SYan Zheng if (!next || next->level <= node->level) 19385d4f98a2SYan Zheng break; 19395d4f98a2SYan Zheng } 19405d4f98a2SYan Zheng 19415d4f98a2SYan Zheng if (!root && !loop && !reloc_only) { 19425d4f98a2SYan Zheng loop = 1; 19435d4f98a2SYan Zheng goto again; 19445d4f98a2SYan Zheng } 19455d4f98a2SYan Zheng 19465d4f98a2SYan Zheng if (root) 19475d4f98a2SYan Zheng *nr = index; 19485d4f98a2SYan Zheng else 19495d4f98a2SYan Zheng *nr = 0; 19505d4f98a2SYan Zheng 19515d4f98a2SYan Zheng return root; 19525d4f98a2SYan Zheng } 19535d4f98a2SYan Zheng 19545d4f98a2SYan Zheng static noinline_for_stack 19555d4f98a2SYan Zheng struct btrfs_root *select_one_root(struct btrfs_trans_handle *trans, 19565d4f98a2SYan Zheng struct backref_node *node) 19575d4f98a2SYan Zheng { 19585d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 19595d4f98a2SYan Zheng int nr; 19605d4f98a2SYan Zheng return __select_one_root(trans, node, edges, &nr, 0); 19615d4f98a2SYan Zheng } 19625d4f98a2SYan Zheng 19635d4f98a2SYan Zheng static noinline_for_stack 19645d4f98a2SYan Zheng struct btrfs_root *select_reloc_root(struct btrfs_trans_handle *trans, 19655d4f98a2SYan Zheng struct backref_node *node, 19665d4f98a2SYan Zheng struct backref_edge *edges[], int *nr) 19675d4f98a2SYan Zheng { 19685d4f98a2SYan Zheng return __select_one_root(trans, node, edges, nr, 1); 19695d4f98a2SYan Zheng } 19705d4f98a2SYan Zheng 19715d4f98a2SYan Zheng static void grab_path_buffers(struct btrfs_path *path, 19725d4f98a2SYan Zheng struct backref_node *node, 19735d4f98a2SYan Zheng struct backref_edge *edges[], int nr) 19745d4f98a2SYan Zheng { 19755d4f98a2SYan Zheng int i = 0; 19765d4f98a2SYan Zheng while (1) { 19775d4f98a2SYan Zheng drop_node_buffer(node); 19785d4f98a2SYan Zheng node->eb = path->nodes[node->level]; 19795d4f98a2SYan Zheng BUG_ON(!node->eb); 19805d4f98a2SYan Zheng if (path->locks[node->level]) 19815d4f98a2SYan Zheng node->locked = 1; 19825d4f98a2SYan Zheng path->nodes[node->level] = NULL; 19835d4f98a2SYan Zheng path->locks[node->level] = 0; 19845d4f98a2SYan Zheng 19855d4f98a2SYan Zheng if (i >= nr) 19865d4f98a2SYan Zheng break; 19875d4f98a2SYan Zheng 19885d4f98a2SYan Zheng edges[i]->blockptr = node->eb->start; 19895d4f98a2SYan Zheng node = edges[i]->node[UPPER]; 19905d4f98a2SYan Zheng i++; 19915d4f98a2SYan Zheng } 19925d4f98a2SYan Zheng } 19935d4f98a2SYan Zheng 19945d4f98a2SYan Zheng /* 19955d4f98a2SYan Zheng * relocate a block tree, and then update pointers in upper level 19965d4f98a2SYan Zheng * blocks that reference the block to point to the new location. 19975d4f98a2SYan Zheng * 19985d4f98a2SYan Zheng * if called by link_to_upper, the block has already been relocated. 19995d4f98a2SYan Zheng * in that case this function just updates pointers. 20005d4f98a2SYan Zheng */ 20015d4f98a2SYan Zheng static int do_relocation(struct btrfs_trans_handle *trans, 20025d4f98a2SYan Zheng struct backref_node *node, 20035d4f98a2SYan Zheng struct btrfs_key *key, 20045d4f98a2SYan Zheng struct btrfs_path *path, int lowest) 20055d4f98a2SYan Zheng { 20065d4f98a2SYan Zheng struct backref_node *upper; 20075d4f98a2SYan Zheng struct backref_edge *edge; 20085d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 20095d4f98a2SYan Zheng struct btrfs_root *root; 20105d4f98a2SYan Zheng struct extent_buffer *eb; 20115d4f98a2SYan Zheng u32 blocksize; 20125d4f98a2SYan Zheng u64 bytenr; 20135d4f98a2SYan Zheng u64 generation; 20145d4f98a2SYan Zheng int nr; 20155d4f98a2SYan Zheng int slot; 20165d4f98a2SYan Zheng int ret; 20175d4f98a2SYan Zheng int err = 0; 20185d4f98a2SYan Zheng 20195d4f98a2SYan Zheng BUG_ON(lowest && node->eb); 20205d4f98a2SYan Zheng 20215d4f98a2SYan Zheng path->lowest_level = node->level + 1; 20225d4f98a2SYan Zheng list_for_each_entry(edge, &node->upper, list[LOWER]) { 20235d4f98a2SYan Zheng cond_resched(); 20245d4f98a2SYan Zheng if (node->eb && node->eb->start == edge->blockptr) 20255d4f98a2SYan Zheng continue; 20265d4f98a2SYan Zheng 20275d4f98a2SYan Zheng upper = edge->node[UPPER]; 20285d4f98a2SYan Zheng root = select_reloc_root(trans, upper, edges, &nr); 20295d4f98a2SYan Zheng if (!root) 20305d4f98a2SYan Zheng continue; 20315d4f98a2SYan Zheng 20325d4f98a2SYan Zheng if (upper->eb && !upper->locked) 20335d4f98a2SYan Zheng drop_node_buffer(upper); 20345d4f98a2SYan Zheng 20355d4f98a2SYan Zheng if (!upper->eb) { 20365d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 20375d4f98a2SYan Zheng if (ret < 0) { 20385d4f98a2SYan Zheng err = ret; 20395d4f98a2SYan Zheng break; 20405d4f98a2SYan Zheng } 20415d4f98a2SYan Zheng BUG_ON(ret > 0); 20425d4f98a2SYan Zheng 20435d4f98a2SYan Zheng slot = path->slots[upper->level]; 20445d4f98a2SYan Zheng 20455d4f98a2SYan Zheng btrfs_unlock_up_safe(path, upper->level + 1); 20465d4f98a2SYan Zheng grab_path_buffers(path, upper, edges, nr); 20475d4f98a2SYan Zheng 20485d4f98a2SYan Zheng btrfs_release_path(NULL, path); 20495d4f98a2SYan Zheng } else { 20505d4f98a2SYan Zheng ret = btrfs_bin_search(upper->eb, key, upper->level, 20515d4f98a2SYan Zheng &slot); 20525d4f98a2SYan Zheng BUG_ON(ret); 20535d4f98a2SYan Zheng } 20545d4f98a2SYan Zheng 20555d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(upper->eb, slot); 20565d4f98a2SYan Zheng if (!lowest) { 20575d4f98a2SYan Zheng if (node->eb->start == bytenr) { 20585d4f98a2SYan Zheng btrfs_tree_unlock(upper->eb); 20595d4f98a2SYan Zheng upper->locked = 0; 20605d4f98a2SYan Zheng continue; 20615d4f98a2SYan Zheng } 20625d4f98a2SYan Zheng } else { 20635d4f98a2SYan Zheng BUG_ON(node->bytenr != bytenr); 20645d4f98a2SYan Zheng } 20655d4f98a2SYan Zheng 20665d4f98a2SYan Zheng blocksize = btrfs_level_size(root, node->level); 20675d4f98a2SYan Zheng generation = btrfs_node_ptr_generation(upper->eb, slot); 20685d4f98a2SYan Zheng eb = read_tree_block(root, bytenr, blocksize, generation); 20695d4f98a2SYan Zheng btrfs_tree_lock(eb); 20705d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 20715d4f98a2SYan Zheng 20725d4f98a2SYan Zheng if (!node->eb) { 20735d4f98a2SYan Zheng ret = btrfs_cow_block(trans, root, eb, upper->eb, 20745d4f98a2SYan Zheng slot, &eb); 20755d4f98a2SYan Zheng if (ret < 0) { 20765d4f98a2SYan Zheng err = ret; 20775d4f98a2SYan Zheng break; 20785d4f98a2SYan Zheng } 20795d4f98a2SYan Zheng btrfs_set_lock_blocking(eb); 20805d4f98a2SYan Zheng node->eb = eb; 20815d4f98a2SYan Zheng node->locked = 1; 20825d4f98a2SYan Zheng } else { 20835d4f98a2SYan Zheng btrfs_set_node_blockptr(upper->eb, slot, 20845d4f98a2SYan Zheng node->eb->start); 20855d4f98a2SYan Zheng btrfs_set_node_ptr_generation(upper->eb, slot, 20865d4f98a2SYan Zheng trans->transid); 20875d4f98a2SYan Zheng btrfs_mark_buffer_dirty(upper->eb); 20885d4f98a2SYan Zheng 20895d4f98a2SYan Zheng ret = btrfs_inc_extent_ref(trans, root, 20905d4f98a2SYan Zheng node->eb->start, blocksize, 20915d4f98a2SYan Zheng upper->eb->start, 20925d4f98a2SYan Zheng btrfs_header_owner(upper->eb), 20935d4f98a2SYan Zheng node->level, 0); 20945d4f98a2SYan Zheng BUG_ON(ret); 20955d4f98a2SYan Zheng 20965d4f98a2SYan Zheng ret = btrfs_drop_subtree(trans, root, eb, upper->eb); 20975d4f98a2SYan Zheng BUG_ON(ret); 20985d4f98a2SYan Zheng } 20995d4f98a2SYan Zheng if (!lowest) { 21005d4f98a2SYan Zheng btrfs_tree_unlock(upper->eb); 21015d4f98a2SYan Zheng upper->locked = 0; 21025d4f98a2SYan Zheng } 21035d4f98a2SYan Zheng } 21045d4f98a2SYan Zheng path->lowest_level = 0; 21055d4f98a2SYan Zheng return err; 21065d4f98a2SYan Zheng } 21075d4f98a2SYan Zheng 21085d4f98a2SYan Zheng static int link_to_upper(struct btrfs_trans_handle *trans, 21095d4f98a2SYan Zheng struct backref_node *node, 21105d4f98a2SYan Zheng struct btrfs_path *path) 21115d4f98a2SYan Zheng { 21125d4f98a2SYan Zheng struct btrfs_key key; 21135d4f98a2SYan Zheng if (!node->eb || list_empty(&node->upper)) 21145d4f98a2SYan Zheng return 0; 21155d4f98a2SYan Zheng 21165d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &key, 0); 21175d4f98a2SYan Zheng return do_relocation(trans, node, &key, path, 0); 21185d4f98a2SYan Zheng } 21195d4f98a2SYan Zheng 21205d4f98a2SYan Zheng static int finish_pending_nodes(struct btrfs_trans_handle *trans, 21215d4f98a2SYan Zheng struct backref_cache *cache, 21225d4f98a2SYan Zheng struct btrfs_path *path) 21235d4f98a2SYan Zheng { 21245d4f98a2SYan Zheng struct backref_node *node; 21255d4f98a2SYan Zheng int level; 21265d4f98a2SYan Zheng int ret; 21275d4f98a2SYan Zheng int err = 0; 21285d4f98a2SYan Zheng 21295d4f98a2SYan Zheng for (level = 0; level < BTRFS_MAX_LEVEL; level++) { 21305d4f98a2SYan Zheng while (!list_empty(&cache->pending[level])) { 21315d4f98a2SYan Zheng node = list_entry(cache->pending[level].next, 21325d4f98a2SYan Zheng struct backref_node, lower); 21335d4f98a2SYan Zheng BUG_ON(node->level != level); 21345d4f98a2SYan Zheng 21355d4f98a2SYan Zheng ret = link_to_upper(trans, node, path); 21365d4f98a2SYan Zheng if (ret < 0) 21375d4f98a2SYan Zheng err = ret; 21385d4f98a2SYan Zheng /* 21395d4f98a2SYan Zheng * this remove the node from the pending list and 21405d4f98a2SYan Zheng * may add some other nodes to the level + 1 21415d4f98a2SYan Zheng * pending list 21425d4f98a2SYan Zheng */ 21435d4f98a2SYan Zheng remove_backref_node(cache, node); 21445d4f98a2SYan Zheng } 21455d4f98a2SYan Zheng } 21465d4f98a2SYan Zheng BUG_ON(!RB_EMPTY_ROOT(&cache->rb_root)); 21475d4f98a2SYan Zheng return err; 21485d4f98a2SYan Zheng } 21495d4f98a2SYan Zheng 21505d4f98a2SYan Zheng static void mark_block_processed(struct reloc_control *rc, 21515d4f98a2SYan Zheng struct backref_node *node) 21525d4f98a2SYan Zheng { 21535d4f98a2SYan Zheng u32 blocksize; 21545d4f98a2SYan Zheng if (node->level == 0 || 21555d4f98a2SYan Zheng in_block_group(node->bytenr, rc->block_group)) { 21565d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level); 21575d4f98a2SYan Zheng set_extent_bits(&rc->processed_blocks, node->bytenr, 21585d4f98a2SYan Zheng node->bytenr + blocksize - 1, EXTENT_DIRTY, 21595d4f98a2SYan Zheng GFP_NOFS); 21605d4f98a2SYan Zheng } 21615d4f98a2SYan Zheng node->processed = 1; 21625d4f98a2SYan Zheng } 21635d4f98a2SYan Zheng 21645d4f98a2SYan Zheng /* 21655d4f98a2SYan Zheng * mark a block and all blocks directly/indirectly reference the block 21665d4f98a2SYan Zheng * as processed. 21675d4f98a2SYan Zheng */ 21685d4f98a2SYan Zheng static void update_processed_blocks(struct reloc_control *rc, 21695d4f98a2SYan Zheng struct backref_node *node) 21705d4f98a2SYan Zheng { 21715d4f98a2SYan Zheng struct backref_node *next = node; 21725d4f98a2SYan Zheng struct backref_edge *edge; 21735d4f98a2SYan Zheng struct backref_edge *edges[BTRFS_MAX_LEVEL - 1]; 21745d4f98a2SYan Zheng int index = 0; 21755d4f98a2SYan Zheng 21765d4f98a2SYan Zheng while (next) { 21775d4f98a2SYan Zheng cond_resched(); 21785d4f98a2SYan Zheng while (1) { 21795d4f98a2SYan Zheng if (next->processed) 21805d4f98a2SYan Zheng break; 21815d4f98a2SYan Zheng 21825d4f98a2SYan Zheng mark_block_processed(rc, next); 21835d4f98a2SYan Zheng 21845d4f98a2SYan Zheng if (list_empty(&next->upper)) 21855d4f98a2SYan Zheng break; 21865d4f98a2SYan Zheng 21875d4f98a2SYan Zheng edge = list_entry(next->upper.next, 21885d4f98a2SYan Zheng struct backref_edge, list[LOWER]); 21895d4f98a2SYan Zheng edges[index++] = edge; 21905d4f98a2SYan Zheng next = edge->node[UPPER]; 21915d4f98a2SYan Zheng } 21925d4f98a2SYan Zheng next = walk_down_backref(edges, &index); 21935d4f98a2SYan Zheng } 21945d4f98a2SYan Zheng } 21955d4f98a2SYan Zheng 21965d4f98a2SYan Zheng static int tree_block_processed(u64 bytenr, u32 blocksize, 21975d4f98a2SYan Zheng struct reloc_control *rc) 21985d4f98a2SYan Zheng { 21995d4f98a2SYan Zheng if (test_range_bit(&rc->processed_blocks, bytenr, 22009655d298SChris Mason bytenr + blocksize - 1, EXTENT_DIRTY, 1, NULL)) 22015d4f98a2SYan Zheng return 1; 22025d4f98a2SYan Zheng return 0; 22035d4f98a2SYan Zheng } 22045d4f98a2SYan Zheng 22055d4f98a2SYan Zheng /* 22065d4f98a2SYan Zheng * check if there are any file extent pointers in the leaf point to 22075d4f98a2SYan Zheng * data require processing 22085d4f98a2SYan Zheng */ 22095d4f98a2SYan Zheng static int check_file_extents(struct reloc_control *rc, 22105d4f98a2SYan Zheng u64 bytenr, u32 blocksize, u64 ptr_gen) 22115d4f98a2SYan Zheng { 22125d4f98a2SYan Zheng struct btrfs_key found_key; 22135d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 22145d4f98a2SYan Zheng struct extent_buffer *leaf; 22155d4f98a2SYan Zheng u32 nritems; 22165d4f98a2SYan Zheng int i; 22175d4f98a2SYan Zheng int ret = 0; 22185d4f98a2SYan Zheng 22195d4f98a2SYan Zheng leaf = read_tree_block(rc->extent_root, bytenr, blocksize, ptr_gen); 22205d4f98a2SYan Zheng 22215d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 22225d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 22235d4f98a2SYan Zheng cond_resched(); 22245d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &found_key, i); 22255d4f98a2SYan Zheng if (found_key.type != BTRFS_EXTENT_DATA_KEY) 22265d4f98a2SYan Zheng continue; 22275d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item); 22285d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 22295d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 22305d4f98a2SYan Zheng continue; 22315d4f98a2SYan Zheng bytenr = btrfs_file_extent_disk_bytenr(leaf, fi); 22325d4f98a2SYan Zheng if (bytenr == 0) 22335d4f98a2SYan Zheng continue; 22345d4f98a2SYan Zheng if (in_block_group(bytenr, rc->block_group)) { 22355d4f98a2SYan Zheng ret = 1; 22365d4f98a2SYan Zheng break; 22375d4f98a2SYan Zheng } 22385d4f98a2SYan Zheng } 22395d4f98a2SYan Zheng free_extent_buffer(leaf); 22405d4f98a2SYan Zheng return ret; 22415d4f98a2SYan Zheng } 22425d4f98a2SYan Zheng 22435d4f98a2SYan Zheng /* 22445d4f98a2SYan Zheng * scan child blocks of a given block to find blocks require processing 22455d4f98a2SYan Zheng */ 22465d4f98a2SYan Zheng static int add_child_blocks(struct btrfs_trans_handle *trans, 22475d4f98a2SYan Zheng struct reloc_control *rc, 22485d4f98a2SYan Zheng struct backref_node *node, 22495d4f98a2SYan Zheng struct rb_root *blocks) 22505d4f98a2SYan Zheng { 22515d4f98a2SYan Zheng struct tree_block *block; 22525d4f98a2SYan Zheng struct rb_node *rb_node; 22535d4f98a2SYan Zheng u64 bytenr; 22545d4f98a2SYan Zheng u64 ptr_gen; 22555d4f98a2SYan Zheng u32 blocksize; 22565d4f98a2SYan Zheng u32 nritems; 22575d4f98a2SYan Zheng int i; 22585d4f98a2SYan Zheng int err = 0; 22595d4f98a2SYan Zheng 22605d4f98a2SYan Zheng nritems = btrfs_header_nritems(node->eb); 22615d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, node->level - 1); 22625d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 22635d4f98a2SYan Zheng cond_resched(); 22645d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(node->eb, i); 22655d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(node->eb, i); 22665d4f98a2SYan Zheng if (ptr_gen == trans->transid) 22675d4f98a2SYan Zheng continue; 22685d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group) && 22695d4f98a2SYan Zheng (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS)) 22705d4f98a2SYan Zheng continue; 22715d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 22725d4f98a2SYan Zheng continue; 22735d4f98a2SYan Zheng 22745d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, 22755d4f98a2SYan Zheng bytenr, blocksize, ptr_gen); 22765d4f98a2SYan Zheng } 22775d4f98a2SYan Zheng 22785d4f98a2SYan Zheng for (i = 0; i < nritems; i++) { 22795d4f98a2SYan Zheng cond_resched(); 22805d4f98a2SYan Zheng bytenr = btrfs_node_blockptr(node->eb, i); 22815d4f98a2SYan Zheng ptr_gen = btrfs_node_ptr_generation(node->eb, i); 22825d4f98a2SYan Zheng if (ptr_gen == trans->transid) 22835d4f98a2SYan Zheng continue; 22845d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group) && 22855d4f98a2SYan Zheng (node->level > 1 || rc->stage == MOVE_DATA_EXTENTS)) 22865d4f98a2SYan Zheng continue; 22875d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 22885d4f98a2SYan Zheng continue; 22895d4f98a2SYan Zheng if (!in_block_group(bytenr, rc->block_group) && 22905d4f98a2SYan Zheng !check_file_extents(rc, bytenr, blocksize, ptr_gen)) 22915d4f98a2SYan Zheng continue; 22925d4f98a2SYan Zheng 22935d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 22945d4f98a2SYan Zheng if (!block) { 22955d4f98a2SYan Zheng err = -ENOMEM; 22965d4f98a2SYan Zheng break; 22975d4f98a2SYan Zheng } 22985d4f98a2SYan Zheng block->bytenr = bytenr; 22995d4f98a2SYan Zheng btrfs_node_key_to_cpu(node->eb, &block->key, i); 23005d4f98a2SYan Zheng block->level = node->level - 1; 23015d4f98a2SYan Zheng block->key_ready = 1; 23025d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 23035d4f98a2SYan Zheng BUG_ON(rb_node); 23045d4f98a2SYan Zheng } 23055d4f98a2SYan Zheng if (err) 23065d4f98a2SYan Zheng free_block_list(blocks); 23075d4f98a2SYan Zheng return err; 23085d4f98a2SYan Zheng } 23095d4f98a2SYan Zheng 23105d4f98a2SYan Zheng /* 23115d4f98a2SYan Zheng * find adjacent blocks require processing 23125d4f98a2SYan Zheng */ 23135d4f98a2SYan Zheng static noinline_for_stack 23145d4f98a2SYan Zheng int add_adjacent_blocks(struct btrfs_trans_handle *trans, 23155d4f98a2SYan Zheng struct reloc_control *rc, 23165d4f98a2SYan Zheng struct backref_cache *cache, 23175d4f98a2SYan Zheng struct rb_root *blocks, int level, 23185d4f98a2SYan Zheng struct backref_node **upper) 23195d4f98a2SYan Zheng { 23205d4f98a2SYan Zheng struct backref_node *node; 23215d4f98a2SYan Zheng int ret = 0; 23225d4f98a2SYan Zheng 23235d4f98a2SYan Zheng WARN_ON(!list_empty(&cache->pending[level])); 23245d4f98a2SYan Zheng 23255d4f98a2SYan Zheng if (list_empty(&cache->pending[level + 1])) 23265d4f98a2SYan Zheng return 1; 23275d4f98a2SYan Zheng 23285d4f98a2SYan Zheng node = list_entry(cache->pending[level + 1].next, 23295d4f98a2SYan Zheng struct backref_node, lower); 23305d4f98a2SYan Zheng if (node->eb) 23315d4f98a2SYan Zheng ret = add_child_blocks(trans, rc, node, blocks); 23325d4f98a2SYan Zheng 23335d4f98a2SYan Zheng *upper = node; 23345d4f98a2SYan Zheng return ret; 23355d4f98a2SYan Zheng } 23365d4f98a2SYan Zheng 23375d4f98a2SYan Zheng static int get_tree_block_key(struct reloc_control *rc, 23385d4f98a2SYan Zheng struct tree_block *block) 23395d4f98a2SYan Zheng { 23405d4f98a2SYan Zheng struct extent_buffer *eb; 23415d4f98a2SYan Zheng 23425d4f98a2SYan Zheng BUG_ON(block->key_ready); 23435d4f98a2SYan Zheng eb = read_tree_block(rc->extent_root, block->bytenr, 23445d4f98a2SYan Zheng block->key.objectid, block->key.offset); 23455d4f98a2SYan Zheng WARN_ON(btrfs_header_level(eb) != block->level); 23465d4f98a2SYan Zheng if (block->level == 0) 23475d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &block->key, 0); 23485d4f98a2SYan Zheng else 23495d4f98a2SYan Zheng btrfs_node_key_to_cpu(eb, &block->key, 0); 23505d4f98a2SYan Zheng free_extent_buffer(eb); 23515d4f98a2SYan Zheng block->key_ready = 1; 23525d4f98a2SYan Zheng return 0; 23535d4f98a2SYan Zheng } 23545d4f98a2SYan Zheng 23555d4f98a2SYan Zheng static int reada_tree_block(struct reloc_control *rc, 23565d4f98a2SYan Zheng struct tree_block *block) 23575d4f98a2SYan Zheng { 23585d4f98a2SYan Zheng BUG_ON(block->key_ready); 23595d4f98a2SYan Zheng readahead_tree_block(rc->extent_root, block->bytenr, 23605d4f98a2SYan Zheng block->key.objectid, block->key.offset); 23615d4f98a2SYan Zheng return 0; 23625d4f98a2SYan Zheng } 23635d4f98a2SYan Zheng 23645d4f98a2SYan Zheng /* 23655d4f98a2SYan Zheng * helper function to relocate a tree block 23665d4f98a2SYan Zheng */ 23675d4f98a2SYan Zheng static int relocate_tree_block(struct btrfs_trans_handle *trans, 23685d4f98a2SYan Zheng struct reloc_control *rc, 23695d4f98a2SYan Zheng struct backref_node *node, 23705d4f98a2SYan Zheng struct btrfs_key *key, 23715d4f98a2SYan Zheng struct btrfs_path *path) 23725d4f98a2SYan Zheng { 23735d4f98a2SYan Zheng struct btrfs_root *root; 23745d4f98a2SYan Zheng int ret; 23755d4f98a2SYan Zheng 23765d4f98a2SYan Zheng root = select_one_root(trans, node); 23775d4f98a2SYan Zheng if (unlikely(!root)) { 23785d4f98a2SYan Zheng rc->found_old_snapshot = 1; 23795d4f98a2SYan Zheng update_processed_blocks(rc, node); 23805d4f98a2SYan Zheng return 0; 23815d4f98a2SYan Zheng } 23825d4f98a2SYan Zheng 23835d4f98a2SYan Zheng if (root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID) { 23845d4f98a2SYan Zheng ret = do_relocation(trans, node, key, path, 1); 23855d4f98a2SYan Zheng if (ret < 0) 23865d4f98a2SYan Zheng goto out; 23875d4f98a2SYan Zheng if (node->level == 0 && rc->stage == UPDATE_DATA_PTRS) { 23885d4f98a2SYan Zheng ret = replace_file_extents(trans, rc, root, 23895d4f98a2SYan Zheng node->eb, NULL); 23905d4f98a2SYan Zheng if (ret < 0) 23915d4f98a2SYan Zheng goto out; 23925d4f98a2SYan Zheng } 23935d4f98a2SYan Zheng drop_node_buffer(node); 23945d4f98a2SYan Zheng } else if (!root->ref_cows) { 23955d4f98a2SYan Zheng path->lowest_level = node->level; 23965d4f98a2SYan Zheng ret = btrfs_search_slot(trans, root, key, path, 0, 1); 23975d4f98a2SYan Zheng btrfs_release_path(root, path); 23985d4f98a2SYan Zheng if (ret < 0) 23995d4f98a2SYan Zheng goto out; 24005d4f98a2SYan Zheng } else if (root != node->root) { 24015d4f98a2SYan Zheng WARN_ON(node->level > 0 || rc->stage != UPDATE_DATA_PTRS); 24025d4f98a2SYan Zheng } 24035d4f98a2SYan Zheng 24045d4f98a2SYan Zheng update_processed_blocks(rc, node); 24055d4f98a2SYan Zheng ret = 0; 24065d4f98a2SYan Zheng out: 24075d4f98a2SYan Zheng drop_node_buffer(node); 24085d4f98a2SYan Zheng return ret; 24095d4f98a2SYan Zheng } 24105d4f98a2SYan Zheng 24115d4f98a2SYan Zheng /* 24125d4f98a2SYan Zheng * relocate a list of blocks 24135d4f98a2SYan Zheng */ 24145d4f98a2SYan Zheng static noinline_for_stack 24155d4f98a2SYan Zheng int relocate_tree_blocks(struct btrfs_trans_handle *trans, 24165d4f98a2SYan Zheng struct reloc_control *rc, struct rb_root *blocks) 24175d4f98a2SYan Zheng { 24185d4f98a2SYan Zheng struct backref_cache *cache; 24195d4f98a2SYan Zheng struct backref_node *node; 24205d4f98a2SYan Zheng struct btrfs_path *path; 24215d4f98a2SYan Zheng struct tree_block *block; 24225d4f98a2SYan Zheng struct rb_node *rb_node; 24235d4f98a2SYan Zheng int level = -1; 24245d4f98a2SYan Zheng int ret; 24255d4f98a2SYan Zheng int err = 0; 24265d4f98a2SYan Zheng 24275d4f98a2SYan Zheng path = btrfs_alloc_path(); 24285d4f98a2SYan Zheng if (!path) 24295d4f98a2SYan Zheng return -ENOMEM; 24305d4f98a2SYan Zheng 24315d4f98a2SYan Zheng cache = kmalloc(sizeof(*cache), GFP_NOFS); 24325d4f98a2SYan Zheng if (!cache) { 24335d4f98a2SYan Zheng btrfs_free_path(path); 24345d4f98a2SYan Zheng return -ENOMEM; 24355d4f98a2SYan Zheng } 24365d4f98a2SYan Zheng 24375d4f98a2SYan Zheng backref_cache_init(cache); 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 if (level == -1) 24435d4f98a2SYan Zheng level = block->level; 24445d4f98a2SYan Zheng else 24455d4f98a2SYan Zheng BUG_ON(level != block->level); 24465d4f98a2SYan Zheng if (!block->key_ready) 24475d4f98a2SYan Zheng reada_tree_block(rc, block); 24485d4f98a2SYan Zheng rb_node = rb_next(rb_node); 24495d4f98a2SYan Zheng } 24505d4f98a2SYan Zheng 24515d4f98a2SYan Zheng rb_node = rb_first(blocks); 24525d4f98a2SYan Zheng while (rb_node) { 24535d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 24545d4f98a2SYan Zheng if (!block->key_ready) 24555d4f98a2SYan Zheng get_tree_block_key(rc, block); 24565d4f98a2SYan Zheng rb_node = rb_next(rb_node); 24575d4f98a2SYan Zheng } 24585d4f98a2SYan Zheng 24595d4f98a2SYan Zheng rb_node = rb_first(blocks); 24605d4f98a2SYan Zheng while (rb_node) { 24615d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 24625d4f98a2SYan Zheng 24635d4f98a2SYan Zheng node = build_backref_tree(rc, cache, &block->key, 24645d4f98a2SYan Zheng block->level, block->bytenr); 24655d4f98a2SYan Zheng if (IS_ERR(node)) { 24665d4f98a2SYan Zheng err = PTR_ERR(node); 24675d4f98a2SYan Zheng goto out; 24685d4f98a2SYan Zheng } 24695d4f98a2SYan Zheng 24705d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, &block->key, 24715d4f98a2SYan Zheng path); 24725d4f98a2SYan Zheng if (ret < 0) { 24735d4f98a2SYan Zheng err = ret; 24745d4f98a2SYan Zheng goto out; 24755d4f98a2SYan Zheng } 24765d4f98a2SYan Zheng remove_backref_node(cache, node); 24775d4f98a2SYan Zheng rb_node = rb_next(rb_node); 24785d4f98a2SYan Zheng } 24795d4f98a2SYan Zheng 24805d4f98a2SYan Zheng if (level > 0) 24815d4f98a2SYan Zheng goto out; 24825d4f98a2SYan Zheng 24835d4f98a2SYan Zheng free_block_list(blocks); 24845d4f98a2SYan Zheng 24855d4f98a2SYan Zheng /* 24865d4f98a2SYan Zheng * now backrefs of some upper level tree blocks have been cached, 24875d4f98a2SYan Zheng * try relocating blocks referenced by these upper level blocks. 24885d4f98a2SYan Zheng */ 24895d4f98a2SYan Zheng while (1) { 24905d4f98a2SYan Zheng struct backref_node *upper = NULL; 24915d4f98a2SYan Zheng if (trans->transaction->in_commit || 24925d4f98a2SYan Zheng trans->transaction->delayed_refs.flushing) 24935d4f98a2SYan Zheng break; 24945d4f98a2SYan Zheng 24955d4f98a2SYan Zheng ret = add_adjacent_blocks(trans, rc, cache, blocks, level, 24965d4f98a2SYan Zheng &upper); 24975d4f98a2SYan Zheng if (ret < 0) 24985d4f98a2SYan Zheng err = ret; 24995d4f98a2SYan Zheng if (ret != 0) 25005d4f98a2SYan Zheng break; 25015d4f98a2SYan Zheng 25025d4f98a2SYan Zheng rb_node = rb_first(blocks); 25035d4f98a2SYan Zheng while (rb_node) { 25045d4f98a2SYan Zheng block = rb_entry(rb_node, struct tree_block, rb_node); 25055d4f98a2SYan Zheng if (trans->transaction->in_commit || 25065d4f98a2SYan Zheng trans->transaction->delayed_refs.flushing) 25075d4f98a2SYan Zheng goto out; 25085d4f98a2SYan Zheng BUG_ON(!block->key_ready); 25095d4f98a2SYan Zheng node = build_backref_tree(rc, cache, &block->key, 25105d4f98a2SYan Zheng level, block->bytenr); 25115d4f98a2SYan Zheng if (IS_ERR(node)) { 25125d4f98a2SYan Zheng err = PTR_ERR(node); 25135d4f98a2SYan Zheng goto out; 25145d4f98a2SYan Zheng } 25155d4f98a2SYan Zheng 25165d4f98a2SYan Zheng ret = relocate_tree_block(trans, rc, node, 25175d4f98a2SYan Zheng &block->key, path); 25185d4f98a2SYan Zheng if (ret < 0) { 25195d4f98a2SYan Zheng err = ret; 25205d4f98a2SYan Zheng goto out; 25215d4f98a2SYan Zheng } 25225d4f98a2SYan Zheng remove_backref_node(cache, node); 25235d4f98a2SYan Zheng rb_node = rb_next(rb_node); 25245d4f98a2SYan Zheng } 25255d4f98a2SYan Zheng free_block_list(blocks); 25265d4f98a2SYan Zheng 25275d4f98a2SYan Zheng if (upper) { 25285d4f98a2SYan Zheng ret = link_to_upper(trans, upper, path); 25295d4f98a2SYan Zheng if (ret < 0) { 25305d4f98a2SYan Zheng err = ret; 25315d4f98a2SYan Zheng break; 25325d4f98a2SYan Zheng } 25335d4f98a2SYan Zheng remove_backref_node(cache, upper); 25345d4f98a2SYan Zheng } 25355d4f98a2SYan Zheng } 25365d4f98a2SYan Zheng out: 25375d4f98a2SYan Zheng free_block_list(blocks); 25385d4f98a2SYan Zheng 25395d4f98a2SYan Zheng ret = finish_pending_nodes(trans, cache, path); 25405d4f98a2SYan Zheng if (ret < 0) 25415d4f98a2SYan Zheng err = ret; 25425d4f98a2SYan Zheng 25435d4f98a2SYan Zheng kfree(cache); 25445d4f98a2SYan Zheng btrfs_free_path(path); 25455d4f98a2SYan Zheng return err; 25465d4f98a2SYan Zheng } 25475d4f98a2SYan Zheng 25485d4f98a2SYan Zheng static noinline_for_stack 25490257bb82SYan, Zheng int setup_extent_mapping(struct inode *inode, u64 start, u64 end, 25500257bb82SYan, Zheng u64 block_start) 25515d4f98a2SYan Zheng { 25525d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 25535d4f98a2SYan Zheng struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree; 25545d4f98a2SYan Zheng struct extent_map *em; 25550257bb82SYan, Zheng int ret = 0; 25565d4f98a2SYan Zheng 25575d4f98a2SYan Zheng em = alloc_extent_map(GFP_NOFS); 25580257bb82SYan, Zheng if (!em) 25590257bb82SYan, Zheng return -ENOMEM; 25600257bb82SYan, Zheng 25615d4f98a2SYan Zheng em->start = start; 25620257bb82SYan, Zheng em->len = end + 1 - start; 25630257bb82SYan, Zheng em->block_len = em->len; 25640257bb82SYan, Zheng em->block_start = block_start; 25655d4f98a2SYan Zheng em->bdev = root->fs_info->fs_devices->latest_bdev; 25665d4f98a2SYan Zheng set_bit(EXTENT_FLAG_PINNED, &em->flags); 25675d4f98a2SYan Zheng 25685d4f98a2SYan Zheng lock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 25695d4f98a2SYan Zheng while (1) { 2570890871beSChris Mason write_lock(&em_tree->lock); 25715d4f98a2SYan Zheng ret = add_extent_mapping(em_tree, em); 2572890871beSChris Mason write_unlock(&em_tree->lock); 25735d4f98a2SYan Zheng if (ret != -EEXIST) { 25745d4f98a2SYan Zheng free_extent_map(em); 25755d4f98a2SYan Zheng break; 25765d4f98a2SYan Zheng } 25775d4f98a2SYan Zheng btrfs_drop_extent_cache(inode, start, end, 0); 25785d4f98a2SYan Zheng } 25795d4f98a2SYan Zheng unlock_extent(&BTRFS_I(inode)->io_tree, start, end, GFP_NOFS); 25800257bb82SYan, Zheng return ret; 25810257bb82SYan, Zheng } 25825d4f98a2SYan Zheng 25830257bb82SYan, Zheng static int relocate_file_extent_cluster(struct inode *inode, 25840257bb82SYan, Zheng struct file_extent_cluster *cluster) 25850257bb82SYan, Zheng { 25860257bb82SYan, Zheng u64 page_start; 25870257bb82SYan, Zheng u64 page_end; 25880257bb82SYan, Zheng u64 offset = BTRFS_I(inode)->index_cnt; 25890257bb82SYan, Zheng unsigned long index; 25900257bb82SYan, Zheng unsigned long last_index; 25910257bb82SYan, Zheng unsigned int dirty_page = 0; 25920257bb82SYan, Zheng struct page *page; 25930257bb82SYan, Zheng struct file_ra_state *ra; 25940257bb82SYan, Zheng int nr = 0; 25950257bb82SYan, Zheng int ret = 0; 25960257bb82SYan, Zheng 25970257bb82SYan, Zheng if (!cluster->nr) 25980257bb82SYan, Zheng return 0; 25990257bb82SYan, Zheng 26000257bb82SYan, Zheng ra = kzalloc(sizeof(*ra), GFP_NOFS); 26010257bb82SYan, Zheng if (!ra) 26020257bb82SYan, Zheng return -ENOMEM; 26030257bb82SYan, Zheng 26040257bb82SYan, Zheng index = (cluster->start - offset) >> PAGE_CACHE_SHIFT; 26050257bb82SYan, Zheng last_index = (cluster->end - offset) >> PAGE_CACHE_SHIFT; 26060257bb82SYan, Zheng 26070257bb82SYan, Zheng mutex_lock(&inode->i_mutex); 26080257bb82SYan, Zheng 26090257bb82SYan, Zheng i_size_write(inode, cluster->end + 1 - offset); 26100257bb82SYan, Zheng ret = setup_extent_mapping(inode, cluster->start - offset, 26110257bb82SYan, Zheng cluster->end - offset, cluster->start); 26120257bb82SYan, Zheng if (ret) 26130257bb82SYan, Zheng goto out_unlock; 26140257bb82SYan, Zheng 26150257bb82SYan, Zheng file_ra_state_init(ra, inode->i_mapping); 26160257bb82SYan, Zheng 26170257bb82SYan, Zheng WARN_ON(cluster->start != cluster->boundary[0]); 26180257bb82SYan, Zheng while (index <= last_index) { 26190257bb82SYan, Zheng page = find_lock_page(inode->i_mapping, index); 26200257bb82SYan, Zheng if (!page) { 26210257bb82SYan, Zheng page_cache_sync_readahead(inode->i_mapping, 26220257bb82SYan, Zheng ra, NULL, index, 26230257bb82SYan, Zheng last_index + 1 - index); 26240257bb82SYan, Zheng page = grab_cache_page(inode->i_mapping, index); 26250257bb82SYan, Zheng if (!page) { 26260257bb82SYan, Zheng ret = -ENOMEM; 26270257bb82SYan, Zheng goto out_unlock; 26280257bb82SYan, Zheng } 26290257bb82SYan, Zheng } 26300257bb82SYan, Zheng 26310257bb82SYan, Zheng if (PageReadahead(page)) { 26320257bb82SYan, Zheng page_cache_async_readahead(inode->i_mapping, 26330257bb82SYan, Zheng ra, NULL, page, index, 26340257bb82SYan, Zheng last_index + 1 - index); 26350257bb82SYan, Zheng } 26360257bb82SYan, Zheng 26370257bb82SYan, Zheng if (!PageUptodate(page)) { 26380257bb82SYan, Zheng btrfs_readpage(NULL, page); 26390257bb82SYan, Zheng lock_page(page); 26400257bb82SYan, Zheng if (!PageUptodate(page)) { 26410257bb82SYan, Zheng unlock_page(page); 26420257bb82SYan, Zheng page_cache_release(page); 26430257bb82SYan, Zheng ret = -EIO; 26440257bb82SYan, Zheng goto out_unlock; 26450257bb82SYan, Zheng } 26460257bb82SYan, Zheng } 26470257bb82SYan, Zheng 26480257bb82SYan, Zheng page_start = (u64)page->index << PAGE_CACHE_SHIFT; 26490257bb82SYan, Zheng page_end = page_start + PAGE_CACHE_SIZE - 1; 26500257bb82SYan, Zheng 26510257bb82SYan, Zheng lock_extent(&BTRFS_I(inode)->io_tree, 26520257bb82SYan, Zheng page_start, page_end, GFP_NOFS); 26530257bb82SYan, Zheng 26540257bb82SYan, Zheng set_page_extent_mapped(page); 26550257bb82SYan, Zheng 26560257bb82SYan, Zheng if (nr < cluster->nr && 26570257bb82SYan, Zheng page_start + offset == cluster->boundary[nr]) { 26580257bb82SYan, Zheng set_extent_bits(&BTRFS_I(inode)->io_tree, 26590257bb82SYan, Zheng page_start, page_end, 26600257bb82SYan, Zheng EXTENT_BOUNDARY, GFP_NOFS); 26610257bb82SYan, Zheng nr++; 26620257bb82SYan, Zheng } 26632ac55d41SJosef Bacik btrfs_set_extent_delalloc(inode, page_start, page_end, NULL); 26640257bb82SYan, Zheng 26650257bb82SYan, Zheng set_page_dirty(page); 26660257bb82SYan, Zheng dirty_page++; 26670257bb82SYan, Zheng 26680257bb82SYan, Zheng unlock_extent(&BTRFS_I(inode)->io_tree, 26690257bb82SYan, Zheng page_start, page_end, GFP_NOFS); 26700257bb82SYan, Zheng unlock_page(page); 26710257bb82SYan, Zheng page_cache_release(page); 26720257bb82SYan, Zheng 26730257bb82SYan, Zheng index++; 26740257bb82SYan, Zheng if (nr < cluster->nr && 26750257bb82SYan, Zheng page_end + 1 + offset == cluster->boundary[nr]) { 26760257bb82SYan, Zheng balance_dirty_pages_ratelimited_nr(inode->i_mapping, 26770257bb82SYan, Zheng dirty_page); 26780257bb82SYan, Zheng dirty_page = 0; 26790257bb82SYan, Zheng } 26800257bb82SYan, Zheng } 26810257bb82SYan, Zheng if (dirty_page) { 26820257bb82SYan, Zheng balance_dirty_pages_ratelimited_nr(inode->i_mapping, 26830257bb82SYan, Zheng dirty_page); 26840257bb82SYan, Zheng } 26850257bb82SYan, Zheng WARN_ON(nr != cluster->nr); 26860257bb82SYan, Zheng out_unlock: 26870257bb82SYan, Zheng mutex_unlock(&inode->i_mutex); 26880257bb82SYan, Zheng kfree(ra); 26890257bb82SYan, Zheng return ret; 26900257bb82SYan, Zheng } 26910257bb82SYan, Zheng 26920257bb82SYan, Zheng static noinline_for_stack 26930257bb82SYan, Zheng int relocate_data_extent(struct inode *inode, struct btrfs_key *extent_key, 26940257bb82SYan, Zheng struct file_extent_cluster *cluster) 26950257bb82SYan, Zheng { 26960257bb82SYan, Zheng int ret; 26970257bb82SYan, Zheng 26980257bb82SYan, Zheng if (cluster->nr > 0 && extent_key->objectid != cluster->end + 1) { 26990257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 27000257bb82SYan, Zheng if (ret) 27010257bb82SYan, Zheng return ret; 27020257bb82SYan, Zheng cluster->nr = 0; 27030257bb82SYan, Zheng } 27040257bb82SYan, Zheng 27050257bb82SYan, Zheng if (!cluster->nr) 27060257bb82SYan, Zheng cluster->start = extent_key->objectid; 27070257bb82SYan, Zheng else 27080257bb82SYan, Zheng BUG_ON(cluster->nr >= MAX_EXTENTS); 27090257bb82SYan, Zheng cluster->end = extent_key->objectid + extent_key->offset - 1; 27100257bb82SYan, Zheng cluster->boundary[cluster->nr] = extent_key->objectid; 27110257bb82SYan, Zheng cluster->nr++; 27120257bb82SYan, Zheng 27130257bb82SYan, Zheng if (cluster->nr >= MAX_EXTENTS) { 27140257bb82SYan, Zheng ret = relocate_file_extent_cluster(inode, cluster); 27150257bb82SYan, Zheng if (ret) 27160257bb82SYan, Zheng return ret; 27170257bb82SYan, Zheng cluster->nr = 0; 27180257bb82SYan, Zheng } 27190257bb82SYan, Zheng return 0; 27205d4f98a2SYan Zheng } 27215d4f98a2SYan Zheng 27225d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 27235d4f98a2SYan Zheng static int get_ref_objectid_v0(struct reloc_control *rc, 27245d4f98a2SYan Zheng struct btrfs_path *path, 27255d4f98a2SYan Zheng struct btrfs_key *extent_key, 27265d4f98a2SYan Zheng u64 *ref_objectid, int *path_change) 27275d4f98a2SYan Zheng { 27285d4f98a2SYan Zheng struct btrfs_key key; 27295d4f98a2SYan Zheng struct extent_buffer *leaf; 27305d4f98a2SYan Zheng struct btrfs_extent_ref_v0 *ref0; 27315d4f98a2SYan Zheng int ret; 27325d4f98a2SYan Zheng int slot; 27335d4f98a2SYan Zheng 27345d4f98a2SYan Zheng leaf = path->nodes[0]; 27355d4f98a2SYan Zheng slot = path->slots[0]; 27365d4f98a2SYan Zheng while (1) { 27375d4f98a2SYan Zheng if (slot >= btrfs_header_nritems(leaf)) { 27385d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 27395d4f98a2SYan Zheng if (ret < 0) 27405d4f98a2SYan Zheng return ret; 27415d4f98a2SYan Zheng BUG_ON(ret > 0); 27425d4f98a2SYan Zheng leaf = path->nodes[0]; 27435d4f98a2SYan Zheng slot = path->slots[0]; 27445d4f98a2SYan Zheng if (path_change) 27455d4f98a2SYan Zheng *path_change = 1; 27465d4f98a2SYan Zheng } 27475d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, slot); 27485d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 27495d4f98a2SYan Zheng return -ENOENT; 27505d4f98a2SYan Zheng 27515d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_REF_V0_KEY) { 27525d4f98a2SYan Zheng slot++; 27535d4f98a2SYan Zheng continue; 27545d4f98a2SYan Zheng } 27555d4f98a2SYan Zheng ref0 = btrfs_item_ptr(leaf, slot, 27565d4f98a2SYan Zheng struct btrfs_extent_ref_v0); 27575d4f98a2SYan Zheng *ref_objectid = btrfs_ref_objectid_v0(leaf, ref0); 27585d4f98a2SYan Zheng break; 27595d4f98a2SYan Zheng } 27605d4f98a2SYan Zheng return 0; 27615d4f98a2SYan Zheng } 27625d4f98a2SYan Zheng #endif 27635d4f98a2SYan Zheng 27645d4f98a2SYan Zheng /* 27655d4f98a2SYan Zheng * helper to add a tree block to the list. 27665d4f98a2SYan Zheng * the major work is getting the generation and level of the block 27675d4f98a2SYan Zheng */ 27685d4f98a2SYan Zheng static int add_tree_block(struct reloc_control *rc, 27695d4f98a2SYan Zheng struct btrfs_key *extent_key, 27705d4f98a2SYan Zheng struct btrfs_path *path, 27715d4f98a2SYan Zheng struct rb_root *blocks) 27725d4f98a2SYan Zheng { 27735d4f98a2SYan Zheng struct extent_buffer *eb; 27745d4f98a2SYan Zheng struct btrfs_extent_item *ei; 27755d4f98a2SYan Zheng struct btrfs_tree_block_info *bi; 27765d4f98a2SYan Zheng struct tree_block *block; 27775d4f98a2SYan Zheng struct rb_node *rb_node; 27785d4f98a2SYan Zheng u32 item_size; 27795d4f98a2SYan Zheng int level = -1; 27805d4f98a2SYan Zheng int generation; 27815d4f98a2SYan Zheng 27825d4f98a2SYan Zheng eb = path->nodes[0]; 27835d4f98a2SYan Zheng item_size = btrfs_item_size_nr(eb, path->slots[0]); 27845d4f98a2SYan Zheng 27855d4f98a2SYan Zheng if (item_size >= sizeof(*ei) + sizeof(*bi)) { 27865d4f98a2SYan Zheng ei = btrfs_item_ptr(eb, path->slots[0], 27875d4f98a2SYan Zheng struct btrfs_extent_item); 27885d4f98a2SYan Zheng bi = (struct btrfs_tree_block_info *)(ei + 1); 27895d4f98a2SYan Zheng generation = btrfs_extent_generation(eb, ei); 27905d4f98a2SYan Zheng level = btrfs_tree_block_level(eb, bi); 27915d4f98a2SYan Zheng } else { 27925d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 27935d4f98a2SYan Zheng u64 ref_owner; 27945d4f98a2SYan Zheng int ret; 27955d4f98a2SYan Zheng 27965d4f98a2SYan Zheng BUG_ON(item_size != sizeof(struct btrfs_extent_item_v0)); 27975d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, extent_key, 27985d4f98a2SYan Zheng &ref_owner, NULL); 27995d4f98a2SYan Zheng BUG_ON(ref_owner >= BTRFS_MAX_LEVEL); 28005d4f98a2SYan Zheng level = (int)ref_owner; 28015d4f98a2SYan Zheng /* FIXME: get real generation */ 28025d4f98a2SYan Zheng generation = 0; 28035d4f98a2SYan Zheng #else 28045d4f98a2SYan Zheng BUG(); 28055d4f98a2SYan Zheng #endif 28065d4f98a2SYan Zheng } 28075d4f98a2SYan Zheng 28085d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 28095d4f98a2SYan Zheng 28105d4f98a2SYan Zheng BUG_ON(level == -1); 28115d4f98a2SYan Zheng 28125d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 28135d4f98a2SYan Zheng if (!block) 28145d4f98a2SYan Zheng return -ENOMEM; 28155d4f98a2SYan Zheng 28165d4f98a2SYan Zheng block->bytenr = extent_key->objectid; 28175d4f98a2SYan Zheng block->key.objectid = extent_key->offset; 28185d4f98a2SYan Zheng block->key.offset = generation; 28195d4f98a2SYan Zheng block->level = level; 28205d4f98a2SYan Zheng block->key_ready = 0; 28215d4f98a2SYan Zheng 28225d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, &block->rb_node); 28235d4f98a2SYan Zheng BUG_ON(rb_node); 28245d4f98a2SYan Zheng 28255d4f98a2SYan Zheng return 0; 28265d4f98a2SYan Zheng } 28275d4f98a2SYan Zheng 28285d4f98a2SYan Zheng /* 28295d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_SHARED_DATA_REF_KEY 28305d4f98a2SYan Zheng */ 28315d4f98a2SYan Zheng static int __add_tree_block(struct reloc_control *rc, 28325d4f98a2SYan Zheng u64 bytenr, u32 blocksize, 28335d4f98a2SYan Zheng struct rb_root *blocks) 28345d4f98a2SYan Zheng { 28355d4f98a2SYan Zheng struct btrfs_path *path; 28365d4f98a2SYan Zheng struct btrfs_key key; 28375d4f98a2SYan Zheng int ret; 28385d4f98a2SYan Zheng 28395d4f98a2SYan Zheng if (tree_block_processed(bytenr, blocksize, rc)) 28405d4f98a2SYan Zheng return 0; 28415d4f98a2SYan Zheng 28425d4f98a2SYan Zheng if (tree_search(blocks, bytenr)) 28435d4f98a2SYan Zheng return 0; 28445d4f98a2SYan Zheng 28455d4f98a2SYan Zheng path = btrfs_alloc_path(); 28465d4f98a2SYan Zheng if (!path) 28475d4f98a2SYan Zheng return -ENOMEM; 28485d4f98a2SYan Zheng 28495d4f98a2SYan Zheng key.objectid = bytenr; 28505d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 28515d4f98a2SYan Zheng key.offset = blocksize; 28525d4f98a2SYan Zheng 28535d4f98a2SYan Zheng path->search_commit_root = 1; 28545d4f98a2SYan Zheng path->skip_locking = 1; 28555d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 0, 0); 28565d4f98a2SYan Zheng if (ret < 0) 28575d4f98a2SYan Zheng goto out; 28585d4f98a2SYan Zheng BUG_ON(ret); 28595d4f98a2SYan Zheng 28605d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 28615d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, blocks); 28625d4f98a2SYan Zheng out: 28635d4f98a2SYan Zheng btrfs_free_path(path); 28645d4f98a2SYan Zheng return ret; 28655d4f98a2SYan Zheng } 28665d4f98a2SYan Zheng 28675d4f98a2SYan Zheng /* 28685d4f98a2SYan Zheng * helper to check if the block use full backrefs for pointers in it 28695d4f98a2SYan Zheng */ 28705d4f98a2SYan Zheng static int block_use_full_backref(struct reloc_control *rc, 28715d4f98a2SYan Zheng struct extent_buffer *eb) 28725d4f98a2SYan Zheng { 28735d4f98a2SYan Zheng struct btrfs_path *path; 28745d4f98a2SYan Zheng struct btrfs_extent_item *ei; 28755d4f98a2SYan Zheng struct btrfs_key key; 28765d4f98a2SYan Zheng u64 flags; 28775d4f98a2SYan Zheng int ret; 28785d4f98a2SYan Zheng 28795d4f98a2SYan Zheng if (btrfs_header_flag(eb, BTRFS_HEADER_FLAG_RELOC) || 28805d4f98a2SYan Zheng btrfs_header_backref_rev(eb) < BTRFS_MIXED_BACKREF_REV) 28815d4f98a2SYan Zheng return 1; 28825d4f98a2SYan Zheng 28835d4f98a2SYan Zheng path = btrfs_alloc_path(); 28845d4f98a2SYan Zheng BUG_ON(!path); 28855d4f98a2SYan Zheng 28865d4f98a2SYan Zheng key.objectid = eb->start; 28875d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 28885d4f98a2SYan Zheng key.offset = eb->len; 28895d4f98a2SYan Zheng 28905d4f98a2SYan Zheng path->search_commit_root = 1; 28915d4f98a2SYan Zheng path->skip_locking = 1; 28925d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 28935d4f98a2SYan Zheng &key, path, 0, 0); 28945d4f98a2SYan Zheng BUG_ON(ret); 28955d4f98a2SYan Zheng 28965d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 28975d4f98a2SYan Zheng struct btrfs_extent_item); 28985d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 28995d4f98a2SYan Zheng BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)); 29005d4f98a2SYan Zheng if (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF) 29015d4f98a2SYan Zheng ret = 1; 29025d4f98a2SYan Zheng else 29035d4f98a2SYan Zheng ret = 0; 29045d4f98a2SYan Zheng btrfs_free_path(path); 29055d4f98a2SYan Zheng return ret; 29065d4f98a2SYan Zheng } 29075d4f98a2SYan Zheng 29085d4f98a2SYan Zheng /* 29095d4f98a2SYan Zheng * helper to add tree blocks for backref of type BTRFS_EXTENT_DATA_REF_KEY 29105d4f98a2SYan Zheng * this function scans fs tree to find blocks reference the data extent 29115d4f98a2SYan Zheng */ 29125d4f98a2SYan Zheng static int find_data_references(struct reloc_control *rc, 29135d4f98a2SYan Zheng struct btrfs_key *extent_key, 29145d4f98a2SYan Zheng struct extent_buffer *leaf, 29155d4f98a2SYan Zheng struct btrfs_extent_data_ref *ref, 29165d4f98a2SYan Zheng struct rb_root *blocks) 29175d4f98a2SYan Zheng { 29185d4f98a2SYan Zheng struct btrfs_path *path; 29195d4f98a2SYan Zheng struct tree_block *block; 29205d4f98a2SYan Zheng struct btrfs_root *root; 29215d4f98a2SYan Zheng struct btrfs_file_extent_item *fi; 29225d4f98a2SYan Zheng struct rb_node *rb_node; 29235d4f98a2SYan Zheng struct btrfs_key key; 29245d4f98a2SYan Zheng u64 ref_root; 29255d4f98a2SYan Zheng u64 ref_objectid; 29265d4f98a2SYan Zheng u64 ref_offset; 29275d4f98a2SYan Zheng u32 ref_count; 29285d4f98a2SYan Zheng u32 nritems; 29295d4f98a2SYan Zheng int err = 0; 29305d4f98a2SYan Zheng int added = 0; 29315d4f98a2SYan Zheng int counted; 29325d4f98a2SYan Zheng int ret; 29335d4f98a2SYan Zheng 29345d4f98a2SYan Zheng path = btrfs_alloc_path(); 29355d4f98a2SYan Zheng if (!path) 29365d4f98a2SYan Zheng return -ENOMEM; 29375d4f98a2SYan Zheng 29385d4f98a2SYan Zheng ref_root = btrfs_extent_data_ref_root(leaf, ref); 29395d4f98a2SYan Zheng ref_objectid = btrfs_extent_data_ref_objectid(leaf, ref); 29405d4f98a2SYan Zheng ref_offset = btrfs_extent_data_ref_offset(leaf, ref); 29415d4f98a2SYan Zheng ref_count = btrfs_extent_data_ref_count(leaf, ref); 29425d4f98a2SYan Zheng 29435d4f98a2SYan Zheng root = read_fs_root(rc->extent_root->fs_info, ref_root); 29445d4f98a2SYan Zheng if (IS_ERR(root)) { 29455d4f98a2SYan Zheng err = PTR_ERR(root); 29465d4f98a2SYan Zheng goto out; 29475d4f98a2SYan Zheng } 29485d4f98a2SYan Zheng 29495d4f98a2SYan Zheng key.objectid = ref_objectid; 29505d4f98a2SYan Zheng key.offset = ref_offset; 29515d4f98a2SYan Zheng key.type = BTRFS_EXTENT_DATA_KEY; 29525d4f98a2SYan Zheng 29535d4f98a2SYan Zheng path->search_commit_root = 1; 29545d4f98a2SYan Zheng path->skip_locking = 1; 29555d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 29565d4f98a2SYan Zheng if (ret < 0) { 29575d4f98a2SYan Zheng err = ret; 29585d4f98a2SYan Zheng goto out; 29595d4f98a2SYan Zheng } 29605d4f98a2SYan Zheng 29615d4f98a2SYan Zheng leaf = path->nodes[0]; 29625d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 29635d4f98a2SYan Zheng /* 29645d4f98a2SYan Zheng * the references in tree blocks that use full backrefs 29655d4f98a2SYan Zheng * are not counted in 29665d4f98a2SYan Zheng */ 29675d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 29685d4f98a2SYan Zheng counted = 0; 29695d4f98a2SYan Zheng else 29705d4f98a2SYan Zheng counted = 1; 29715d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 29725d4f98a2SYan Zheng if (rb_node) { 29735d4f98a2SYan Zheng if (counted) 29745d4f98a2SYan Zheng added = 1; 29755d4f98a2SYan Zheng else 29765d4f98a2SYan Zheng path->slots[0] = nritems; 29775d4f98a2SYan Zheng } 29785d4f98a2SYan Zheng 29795d4f98a2SYan Zheng while (ref_count > 0) { 29805d4f98a2SYan Zheng while (path->slots[0] >= nritems) { 29815d4f98a2SYan Zheng ret = btrfs_next_leaf(root, path); 29825d4f98a2SYan Zheng if (ret < 0) { 29835d4f98a2SYan Zheng err = ret; 29845d4f98a2SYan Zheng goto out; 29855d4f98a2SYan Zheng } 29865d4f98a2SYan Zheng if (ret > 0) { 29875d4f98a2SYan Zheng WARN_ON(1); 29885d4f98a2SYan Zheng goto out; 29895d4f98a2SYan Zheng } 29905d4f98a2SYan Zheng 29915d4f98a2SYan Zheng leaf = path->nodes[0]; 29925d4f98a2SYan Zheng nritems = btrfs_header_nritems(leaf); 29935d4f98a2SYan Zheng added = 0; 29945d4f98a2SYan Zheng 29955d4f98a2SYan Zheng if (block_use_full_backref(rc, leaf)) 29965d4f98a2SYan Zheng counted = 0; 29975d4f98a2SYan Zheng else 29985d4f98a2SYan Zheng counted = 1; 29995d4f98a2SYan Zheng rb_node = tree_search(blocks, leaf->start); 30005d4f98a2SYan Zheng if (rb_node) { 30015d4f98a2SYan Zheng if (counted) 30025d4f98a2SYan Zheng added = 1; 30035d4f98a2SYan Zheng else 30045d4f98a2SYan Zheng path->slots[0] = nritems; 30055d4f98a2SYan Zheng } 30065d4f98a2SYan Zheng } 30075d4f98a2SYan Zheng 30085d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 30095d4f98a2SYan Zheng if (key.objectid != ref_objectid || 30105d4f98a2SYan Zheng key.type != BTRFS_EXTENT_DATA_KEY) { 30115d4f98a2SYan Zheng WARN_ON(1); 30125d4f98a2SYan Zheng break; 30135d4f98a2SYan Zheng } 30145d4f98a2SYan Zheng 30155d4f98a2SYan Zheng fi = btrfs_item_ptr(leaf, path->slots[0], 30165d4f98a2SYan Zheng struct btrfs_file_extent_item); 30175d4f98a2SYan Zheng 30185d4f98a2SYan Zheng if (btrfs_file_extent_type(leaf, fi) == 30195d4f98a2SYan Zheng BTRFS_FILE_EXTENT_INLINE) 30205d4f98a2SYan Zheng goto next; 30215d4f98a2SYan Zheng 30225d4f98a2SYan Zheng if (btrfs_file_extent_disk_bytenr(leaf, fi) != 30235d4f98a2SYan Zheng extent_key->objectid) 30245d4f98a2SYan Zheng goto next; 30255d4f98a2SYan Zheng 30265d4f98a2SYan Zheng key.offset -= btrfs_file_extent_offset(leaf, fi); 30275d4f98a2SYan Zheng if (key.offset != ref_offset) 30285d4f98a2SYan Zheng goto next; 30295d4f98a2SYan Zheng 30305d4f98a2SYan Zheng if (counted) 30315d4f98a2SYan Zheng ref_count--; 30325d4f98a2SYan Zheng if (added) 30335d4f98a2SYan Zheng goto next; 30345d4f98a2SYan Zheng 30355d4f98a2SYan Zheng if (!tree_block_processed(leaf->start, leaf->len, rc)) { 30365d4f98a2SYan Zheng block = kmalloc(sizeof(*block), GFP_NOFS); 30375d4f98a2SYan Zheng if (!block) { 30385d4f98a2SYan Zheng err = -ENOMEM; 30395d4f98a2SYan Zheng break; 30405d4f98a2SYan Zheng } 30415d4f98a2SYan Zheng block->bytenr = leaf->start; 30425d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &block->key, 0); 30435d4f98a2SYan Zheng block->level = 0; 30445d4f98a2SYan Zheng block->key_ready = 1; 30455d4f98a2SYan Zheng rb_node = tree_insert(blocks, block->bytenr, 30465d4f98a2SYan Zheng &block->rb_node); 30475d4f98a2SYan Zheng BUG_ON(rb_node); 30485d4f98a2SYan Zheng } 30495d4f98a2SYan Zheng if (counted) 30505d4f98a2SYan Zheng added = 1; 30515d4f98a2SYan Zheng else 30525d4f98a2SYan Zheng path->slots[0] = nritems; 30535d4f98a2SYan Zheng next: 30545d4f98a2SYan Zheng path->slots[0]++; 30555d4f98a2SYan Zheng 30565d4f98a2SYan Zheng } 30575d4f98a2SYan Zheng out: 30585d4f98a2SYan Zheng btrfs_free_path(path); 30595d4f98a2SYan Zheng return err; 30605d4f98a2SYan Zheng } 30615d4f98a2SYan Zheng 30625d4f98a2SYan Zheng /* 30635d4f98a2SYan Zheng * hepler to find all tree blocks that reference a given data extent 30645d4f98a2SYan Zheng */ 30655d4f98a2SYan Zheng static noinline_for_stack 30665d4f98a2SYan Zheng int add_data_references(struct reloc_control *rc, 30675d4f98a2SYan Zheng struct btrfs_key *extent_key, 30685d4f98a2SYan Zheng struct btrfs_path *path, 30695d4f98a2SYan Zheng struct rb_root *blocks) 30705d4f98a2SYan Zheng { 30715d4f98a2SYan Zheng struct btrfs_key key; 30725d4f98a2SYan Zheng struct extent_buffer *eb; 30735d4f98a2SYan Zheng struct btrfs_extent_data_ref *dref; 30745d4f98a2SYan Zheng struct btrfs_extent_inline_ref *iref; 30755d4f98a2SYan Zheng unsigned long ptr; 30765d4f98a2SYan Zheng unsigned long end; 30775d4f98a2SYan Zheng u32 blocksize; 30785d4f98a2SYan Zheng int ret; 30795d4f98a2SYan Zheng int err = 0; 30805d4f98a2SYan Zheng 30815d4f98a2SYan Zheng ret = get_new_location(rc->data_inode, NULL, extent_key->objectid, 30825d4f98a2SYan Zheng extent_key->offset); 30835d4f98a2SYan Zheng BUG_ON(ret < 0); 30845d4f98a2SYan Zheng if (ret > 0) { 30855d4f98a2SYan Zheng /* the relocated data is fragmented */ 30865d4f98a2SYan Zheng rc->extents_skipped++; 30875d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 30885d4f98a2SYan Zheng return 0; 30895d4f98a2SYan Zheng } 30905d4f98a2SYan Zheng 30915d4f98a2SYan Zheng blocksize = btrfs_level_size(rc->extent_root, 0); 30925d4f98a2SYan Zheng 30935d4f98a2SYan Zheng eb = path->nodes[0]; 30945d4f98a2SYan Zheng ptr = btrfs_item_ptr_offset(eb, path->slots[0]); 30955d4f98a2SYan Zheng end = ptr + btrfs_item_size_nr(eb, path->slots[0]); 30965d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 30975d4f98a2SYan Zheng if (ptr + sizeof(struct btrfs_extent_item_v0) == end) 30985d4f98a2SYan Zheng ptr = end; 30995d4f98a2SYan Zheng else 31005d4f98a2SYan Zheng #endif 31015d4f98a2SYan Zheng ptr += sizeof(struct btrfs_extent_item); 31025d4f98a2SYan Zheng 31035d4f98a2SYan Zheng while (ptr < end) { 31045d4f98a2SYan Zheng iref = (struct btrfs_extent_inline_ref *)ptr; 31055d4f98a2SYan Zheng key.type = btrfs_extent_inline_ref_type(eb, iref); 31065d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 31075d4f98a2SYan Zheng key.offset = btrfs_extent_inline_ref_offset(eb, iref); 31085d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 31095d4f98a2SYan Zheng blocks); 31105d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 31115d4f98a2SYan Zheng dref = (struct btrfs_extent_data_ref *)(&iref->offset); 31125d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 31135d4f98a2SYan Zheng eb, dref, blocks); 31145d4f98a2SYan Zheng } else { 31155d4f98a2SYan Zheng BUG(); 31165d4f98a2SYan Zheng } 31175d4f98a2SYan Zheng ptr += btrfs_extent_inline_ref_size(key.type); 31185d4f98a2SYan Zheng } 31195d4f98a2SYan Zheng WARN_ON(ptr > end); 31205d4f98a2SYan Zheng 31215d4f98a2SYan Zheng while (1) { 31225d4f98a2SYan Zheng cond_resched(); 31235d4f98a2SYan Zheng eb = path->nodes[0]; 31245d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(eb)) { 31255d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 31265d4f98a2SYan Zheng if (ret < 0) { 31275d4f98a2SYan Zheng err = ret; 31285d4f98a2SYan Zheng break; 31295d4f98a2SYan Zheng } 31305d4f98a2SYan Zheng if (ret > 0) 31315d4f98a2SYan Zheng break; 31325d4f98a2SYan Zheng eb = path->nodes[0]; 31335d4f98a2SYan Zheng } 31345d4f98a2SYan Zheng 31355d4f98a2SYan Zheng btrfs_item_key_to_cpu(eb, &key, path->slots[0]); 31365d4f98a2SYan Zheng if (key.objectid != extent_key->objectid) 31375d4f98a2SYan Zheng break; 31385d4f98a2SYan Zheng 31395d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 31405d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY || 31415d4f98a2SYan Zheng key.type == BTRFS_EXTENT_REF_V0_KEY) { 31425d4f98a2SYan Zheng #else 31435d4f98a2SYan Zheng BUG_ON(key.type == BTRFS_EXTENT_REF_V0_KEY); 31445d4f98a2SYan Zheng if (key.type == BTRFS_SHARED_DATA_REF_KEY) { 31455d4f98a2SYan Zheng #endif 31465d4f98a2SYan Zheng ret = __add_tree_block(rc, key.offset, blocksize, 31475d4f98a2SYan Zheng blocks); 31485d4f98a2SYan Zheng } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) { 31495d4f98a2SYan Zheng dref = btrfs_item_ptr(eb, path->slots[0], 31505d4f98a2SYan Zheng struct btrfs_extent_data_ref); 31515d4f98a2SYan Zheng ret = find_data_references(rc, extent_key, 31525d4f98a2SYan Zheng eb, dref, blocks); 31535d4f98a2SYan Zheng } else { 31545d4f98a2SYan Zheng ret = 0; 31555d4f98a2SYan Zheng } 31565d4f98a2SYan Zheng if (ret) { 31575d4f98a2SYan Zheng err = ret; 31585d4f98a2SYan Zheng break; 31595d4f98a2SYan Zheng } 31605d4f98a2SYan Zheng path->slots[0]++; 31615d4f98a2SYan Zheng } 31625d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 31635d4f98a2SYan Zheng if (err) 31645d4f98a2SYan Zheng free_block_list(blocks); 31655d4f98a2SYan Zheng return err; 31665d4f98a2SYan Zheng } 31675d4f98a2SYan Zheng 31685d4f98a2SYan Zheng /* 31695d4f98a2SYan Zheng * hepler to find next unprocessed extent 31705d4f98a2SYan Zheng */ 31715d4f98a2SYan Zheng static noinline_for_stack 31725d4f98a2SYan Zheng int find_next_extent(struct btrfs_trans_handle *trans, 31735d4f98a2SYan Zheng struct reloc_control *rc, struct btrfs_path *path) 31745d4f98a2SYan Zheng { 31755d4f98a2SYan Zheng struct btrfs_key key; 31765d4f98a2SYan Zheng struct extent_buffer *leaf; 31775d4f98a2SYan Zheng u64 start, end, last; 31785d4f98a2SYan Zheng int ret; 31795d4f98a2SYan Zheng 31805d4f98a2SYan Zheng last = rc->block_group->key.objectid + rc->block_group->key.offset; 31815d4f98a2SYan Zheng while (1) { 31825d4f98a2SYan Zheng cond_resched(); 31835d4f98a2SYan Zheng if (rc->search_start >= last) { 31845d4f98a2SYan Zheng ret = 1; 31855d4f98a2SYan Zheng break; 31865d4f98a2SYan Zheng } 31875d4f98a2SYan Zheng 31885d4f98a2SYan Zheng key.objectid = rc->search_start; 31895d4f98a2SYan Zheng key.type = BTRFS_EXTENT_ITEM_KEY; 31905d4f98a2SYan Zheng key.offset = 0; 31915d4f98a2SYan Zheng 31925d4f98a2SYan Zheng path->search_commit_root = 1; 31935d4f98a2SYan Zheng path->skip_locking = 1; 31945d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, &key, path, 31955d4f98a2SYan Zheng 0, 0); 31965d4f98a2SYan Zheng if (ret < 0) 31975d4f98a2SYan Zheng break; 31985d4f98a2SYan Zheng next: 31995d4f98a2SYan Zheng leaf = path->nodes[0]; 32005d4f98a2SYan Zheng if (path->slots[0] >= btrfs_header_nritems(leaf)) { 32015d4f98a2SYan Zheng ret = btrfs_next_leaf(rc->extent_root, path); 32025d4f98a2SYan Zheng if (ret != 0) 32035d4f98a2SYan Zheng break; 32045d4f98a2SYan Zheng leaf = path->nodes[0]; 32055d4f98a2SYan Zheng } 32065d4f98a2SYan Zheng 32075d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 32085d4f98a2SYan Zheng if (key.objectid >= last) { 32095d4f98a2SYan Zheng ret = 1; 32105d4f98a2SYan Zheng break; 32115d4f98a2SYan Zheng } 32125d4f98a2SYan Zheng 32135d4f98a2SYan Zheng if (key.type != BTRFS_EXTENT_ITEM_KEY || 32145d4f98a2SYan Zheng key.objectid + key.offset <= rc->search_start) { 32155d4f98a2SYan Zheng path->slots[0]++; 32165d4f98a2SYan Zheng goto next; 32175d4f98a2SYan Zheng } 32185d4f98a2SYan Zheng 32195d4f98a2SYan Zheng ret = find_first_extent_bit(&rc->processed_blocks, 32205d4f98a2SYan Zheng key.objectid, &start, &end, 32215d4f98a2SYan Zheng EXTENT_DIRTY); 32225d4f98a2SYan Zheng 32235d4f98a2SYan Zheng if (ret == 0 && start <= key.objectid) { 32245d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 32255d4f98a2SYan Zheng rc->search_start = end + 1; 32265d4f98a2SYan Zheng } else { 32275d4f98a2SYan Zheng rc->search_start = key.objectid + key.offset; 32285d4f98a2SYan Zheng return 0; 32295d4f98a2SYan Zheng } 32305d4f98a2SYan Zheng } 32315d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 32325d4f98a2SYan Zheng return ret; 32335d4f98a2SYan Zheng } 32345d4f98a2SYan Zheng 32355d4f98a2SYan Zheng static void set_reloc_control(struct reloc_control *rc) 32365d4f98a2SYan Zheng { 32375d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32385d4f98a2SYan Zheng mutex_lock(&fs_info->trans_mutex); 32395d4f98a2SYan Zheng fs_info->reloc_ctl = rc; 32405d4f98a2SYan Zheng mutex_unlock(&fs_info->trans_mutex); 32415d4f98a2SYan Zheng } 32425d4f98a2SYan Zheng 32435d4f98a2SYan Zheng static void unset_reloc_control(struct reloc_control *rc) 32445d4f98a2SYan Zheng { 32455d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = rc->extent_root->fs_info; 32465d4f98a2SYan Zheng mutex_lock(&fs_info->trans_mutex); 32475d4f98a2SYan Zheng fs_info->reloc_ctl = NULL; 32485d4f98a2SYan Zheng mutex_unlock(&fs_info->trans_mutex); 32495d4f98a2SYan Zheng } 32505d4f98a2SYan Zheng 32515d4f98a2SYan Zheng static int check_extent_flags(u64 flags) 32525d4f98a2SYan Zheng { 32535d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 32545d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 32555d4f98a2SYan Zheng return 1; 32565d4f98a2SYan Zheng if (!(flags & BTRFS_EXTENT_FLAG_DATA) && 32575d4f98a2SYan Zheng !(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) 32585d4f98a2SYan Zheng return 1; 32595d4f98a2SYan Zheng if ((flags & BTRFS_EXTENT_FLAG_DATA) && 32605d4f98a2SYan Zheng (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF)) 32615d4f98a2SYan Zheng return 1; 32625d4f98a2SYan Zheng return 0; 32635d4f98a2SYan Zheng } 32645d4f98a2SYan Zheng 326576dda93cSYan, Zheng 32665d4f98a2SYan Zheng static noinline_for_stack int relocate_block_group(struct reloc_control *rc) 32675d4f98a2SYan Zheng { 32685d4f98a2SYan Zheng struct rb_root blocks = RB_ROOT; 32695d4f98a2SYan Zheng struct btrfs_key key; 32700257bb82SYan, Zheng struct file_extent_cluster *cluster; 32715d4f98a2SYan Zheng struct btrfs_trans_handle *trans = NULL; 32725d4f98a2SYan Zheng struct btrfs_path *path; 32735d4f98a2SYan Zheng struct btrfs_extent_item *ei; 32745d4f98a2SYan Zheng unsigned long nr; 32755d4f98a2SYan Zheng u64 flags; 32765d4f98a2SYan Zheng u32 item_size; 32775d4f98a2SYan Zheng int ret; 32785d4f98a2SYan Zheng int err = 0; 32795d4f98a2SYan Zheng 32800257bb82SYan, Zheng cluster = kzalloc(sizeof(*cluster), GFP_NOFS); 32810257bb82SYan, Zheng if (!cluster) 32820257bb82SYan, Zheng return -ENOMEM; 32830257bb82SYan, Zheng 32845d4f98a2SYan Zheng path = btrfs_alloc_path(); 32852423fdfbSJiri Slaby if (!path) { 32862423fdfbSJiri Slaby kfree(cluster); 32875d4f98a2SYan Zheng return -ENOMEM; 32882423fdfbSJiri Slaby } 32895d4f98a2SYan Zheng 329076dda93cSYan, Zheng rc->extents_found = 0; 329176dda93cSYan, Zheng rc->extents_skipped = 0; 329276dda93cSYan, Zheng 32935d4f98a2SYan Zheng rc->search_start = rc->block_group->key.objectid; 32945d4f98a2SYan Zheng clear_extent_bits(&rc->processed_blocks, 0, (u64)-1, EXTENT_DIRTY, 32955d4f98a2SYan Zheng GFP_NOFS); 32965d4f98a2SYan Zheng 32975d4f98a2SYan Zheng rc->create_reloc_root = 1; 32985d4f98a2SYan Zheng set_reloc_control(rc); 32995d4f98a2SYan Zheng 33005d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 33015d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 33025d4f98a2SYan Zheng 33035d4f98a2SYan Zheng while (1) { 33045d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 33055d4f98a2SYan Zheng 33065d4f98a2SYan Zheng ret = find_next_extent(trans, rc, path); 33075d4f98a2SYan Zheng if (ret < 0) 33085d4f98a2SYan Zheng err = ret; 33095d4f98a2SYan Zheng if (ret != 0) 33105d4f98a2SYan Zheng break; 33115d4f98a2SYan Zheng 33125d4f98a2SYan Zheng rc->extents_found++; 33135d4f98a2SYan Zheng 33145d4f98a2SYan Zheng ei = btrfs_item_ptr(path->nodes[0], path->slots[0], 33155d4f98a2SYan Zheng struct btrfs_extent_item); 33165d4f98a2SYan Zheng btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]); 33175d4f98a2SYan Zheng item_size = btrfs_item_size_nr(path->nodes[0], 33185d4f98a2SYan Zheng path->slots[0]); 33195d4f98a2SYan Zheng if (item_size >= sizeof(*ei)) { 33205d4f98a2SYan Zheng flags = btrfs_extent_flags(path->nodes[0], ei); 33215d4f98a2SYan Zheng ret = check_extent_flags(flags); 33225d4f98a2SYan Zheng BUG_ON(ret); 33235d4f98a2SYan Zheng 33245d4f98a2SYan Zheng } else { 33255d4f98a2SYan Zheng #ifdef BTRFS_COMPAT_EXTENT_TREE_V0 33265d4f98a2SYan Zheng u64 ref_owner; 33275d4f98a2SYan Zheng int path_change = 0; 33285d4f98a2SYan Zheng 33295d4f98a2SYan Zheng BUG_ON(item_size != 33305d4f98a2SYan Zheng sizeof(struct btrfs_extent_item_v0)); 33315d4f98a2SYan Zheng ret = get_ref_objectid_v0(rc, path, &key, &ref_owner, 33325d4f98a2SYan Zheng &path_change); 33335d4f98a2SYan Zheng if (ref_owner < BTRFS_FIRST_FREE_OBJECTID) 33345d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_TREE_BLOCK; 33355d4f98a2SYan Zheng else 33365d4f98a2SYan Zheng flags = BTRFS_EXTENT_FLAG_DATA; 33375d4f98a2SYan Zheng 33385d4f98a2SYan Zheng if (path_change) { 33395d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 33405d4f98a2SYan Zheng 33415d4f98a2SYan Zheng path->search_commit_root = 1; 33425d4f98a2SYan Zheng path->skip_locking = 1; 33435d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, rc->extent_root, 33445d4f98a2SYan Zheng &key, path, 0, 0); 33455d4f98a2SYan Zheng if (ret < 0) { 33465d4f98a2SYan Zheng err = ret; 33475d4f98a2SYan Zheng break; 33485d4f98a2SYan Zheng } 33495d4f98a2SYan Zheng BUG_ON(ret > 0); 33505d4f98a2SYan Zheng } 33515d4f98a2SYan Zheng #else 33525d4f98a2SYan Zheng BUG(); 33535d4f98a2SYan Zheng #endif 33545d4f98a2SYan Zheng } 33555d4f98a2SYan Zheng 33565d4f98a2SYan Zheng if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 33575d4f98a2SYan Zheng ret = add_tree_block(rc, &key, path, &blocks); 33585d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 33595d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 33605d4f98a2SYan Zheng ret = add_data_references(rc, &key, path, &blocks); 33615d4f98a2SYan Zheng } else { 33625d4f98a2SYan Zheng btrfs_release_path(rc->extent_root, path); 33635d4f98a2SYan Zheng ret = 0; 33645d4f98a2SYan Zheng } 33655d4f98a2SYan Zheng if (ret < 0) { 33665d4f98a2SYan Zheng err = 0; 33675d4f98a2SYan Zheng break; 33685d4f98a2SYan Zheng } 33695d4f98a2SYan Zheng 33705d4f98a2SYan Zheng if (!RB_EMPTY_ROOT(&blocks)) { 33715d4f98a2SYan Zheng ret = relocate_tree_blocks(trans, rc, &blocks); 33725d4f98a2SYan Zheng if (ret < 0) { 33735d4f98a2SYan Zheng err = ret; 33745d4f98a2SYan Zheng break; 33755d4f98a2SYan Zheng } 33765d4f98a2SYan Zheng } 33775d4f98a2SYan Zheng 33785d4f98a2SYan Zheng nr = trans->blocks_used; 33790257bb82SYan, Zheng btrfs_end_transaction(trans, rc->extent_root); 33805d4f98a2SYan Zheng trans = NULL; 33815d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 33825d4f98a2SYan Zheng 33835d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && 33845d4f98a2SYan Zheng (flags & BTRFS_EXTENT_FLAG_DATA)) { 33855d4f98a2SYan Zheng rc->found_file_extent = 1; 33860257bb82SYan, Zheng ret = relocate_data_extent(rc->data_inode, 33870257bb82SYan, Zheng &key, cluster); 33885d4f98a2SYan Zheng if (ret < 0) { 33895d4f98a2SYan Zheng err = ret; 33905d4f98a2SYan Zheng break; 33915d4f98a2SYan Zheng } 33925d4f98a2SYan Zheng } 33935d4f98a2SYan Zheng } 33945d4f98a2SYan Zheng btrfs_free_path(path); 33955d4f98a2SYan Zheng 33965d4f98a2SYan Zheng if (trans) { 33975d4f98a2SYan Zheng nr = trans->blocks_used; 33985d4f98a2SYan Zheng btrfs_end_transaction(trans, rc->extent_root); 33995d4f98a2SYan Zheng btrfs_btree_balance_dirty(rc->extent_root, nr); 34005d4f98a2SYan Zheng } 34015d4f98a2SYan Zheng 34020257bb82SYan, Zheng if (!err) { 34030257bb82SYan, Zheng ret = relocate_file_extent_cluster(rc->data_inode, cluster); 34040257bb82SYan, Zheng if (ret < 0) 34050257bb82SYan, Zheng err = ret; 34060257bb82SYan, Zheng } 34070257bb82SYan, Zheng 34080257bb82SYan, Zheng kfree(cluster); 34090257bb82SYan, Zheng 34105d4f98a2SYan Zheng rc->create_reloc_root = 0; 34115d4f98a2SYan Zheng smp_mb(); 34125d4f98a2SYan Zheng 34135d4f98a2SYan Zheng if (rc->extents_found > 0) { 34145d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 34155d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 34165d4f98a2SYan Zheng } 34175d4f98a2SYan Zheng 34185d4f98a2SYan Zheng merge_reloc_roots(rc); 34195d4f98a2SYan Zheng 34205d4f98a2SYan Zheng unset_reloc_control(rc); 34215d4f98a2SYan Zheng 34225d4f98a2SYan Zheng /* get rid of pinned extents */ 34235d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 34245d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 34255d4f98a2SYan Zheng 34265d4f98a2SYan Zheng return err; 34275d4f98a2SYan Zheng } 34285d4f98a2SYan Zheng 34295d4f98a2SYan Zheng static int __insert_orphan_inode(struct btrfs_trans_handle *trans, 34300257bb82SYan, Zheng struct btrfs_root *root, u64 objectid) 34315d4f98a2SYan Zheng { 34325d4f98a2SYan Zheng struct btrfs_path *path; 34335d4f98a2SYan Zheng struct btrfs_inode_item *item; 34345d4f98a2SYan Zheng struct extent_buffer *leaf; 34355d4f98a2SYan Zheng int ret; 34365d4f98a2SYan Zheng 34375d4f98a2SYan Zheng path = btrfs_alloc_path(); 34385d4f98a2SYan Zheng if (!path) 34395d4f98a2SYan Zheng return -ENOMEM; 34405d4f98a2SYan Zheng 34415d4f98a2SYan Zheng ret = btrfs_insert_empty_inode(trans, root, path, objectid); 34425d4f98a2SYan Zheng if (ret) 34435d4f98a2SYan Zheng goto out; 34445d4f98a2SYan Zheng 34455d4f98a2SYan Zheng leaf = path->nodes[0]; 34465d4f98a2SYan Zheng item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_inode_item); 34475d4f98a2SYan Zheng memset_extent_buffer(leaf, 0, (unsigned long)item, sizeof(*item)); 34485d4f98a2SYan Zheng btrfs_set_inode_generation(leaf, item, 1); 34490257bb82SYan, Zheng btrfs_set_inode_size(leaf, item, 0); 34505d4f98a2SYan Zheng btrfs_set_inode_mode(leaf, item, S_IFREG | 0600); 34515d4f98a2SYan Zheng btrfs_set_inode_flags(leaf, item, BTRFS_INODE_NOCOMPRESS); 34525d4f98a2SYan Zheng btrfs_mark_buffer_dirty(leaf); 34535d4f98a2SYan Zheng btrfs_release_path(root, path); 34545d4f98a2SYan Zheng out: 34555d4f98a2SYan Zheng btrfs_free_path(path); 34565d4f98a2SYan Zheng return ret; 34575d4f98a2SYan Zheng } 34585d4f98a2SYan Zheng 34595d4f98a2SYan Zheng /* 34605d4f98a2SYan Zheng * helper to create inode for data relocation. 34615d4f98a2SYan Zheng * the inode is in data relocation tree and its link count is 0 34625d4f98a2SYan Zheng */ 34635d4f98a2SYan Zheng static struct inode *create_reloc_inode(struct btrfs_fs_info *fs_info, 34645d4f98a2SYan Zheng struct btrfs_block_group_cache *group) 34655d4f98a2SYan Zheng { 34665d4f98a2SYan Zheng struct inode *inode = NULL; 34675d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 34685d4f98a2SYan Zheng struct btrfs_root *root; 34695d4f98a2SYan Zheng struct btrfs_key key; 34705d4f98a2SYan Zheng unsigned long nr; 34715d4f98a2SYan Zheng u64 objectid = BTRFS_FIRST_FREE_OBJECTID; 34725d4f98a2SYan Zheng int err = 0; 34735d4f98a2SYan Zheng 34745d4f98a2SYan Zheng root = read_fs_root(fs_info, BTRFS_DATA_RELOC_TREE_OBJECTID); 34755d4f98a2SYan Zheng if (IS_ERR(root)) 34765d4f98a2SYan Zheng return ERR_CAST(root); 34775d4f98a2SYan Zheng 34785d4f98a2SYan Zheng trans = btrfs_start_transaction(root, 1); 34795d4f98a2SYan Zheng BUG_ON(!trans); 34805d4f98a2SYan Zheng 34815d4f98a2SYan Zheng err = btrfs_find_free_objectid(trans, root, objectid, &objectid); 34825d4f98a2SYan Zheng if (err) 34835d4f98a2SYan Zheng goto out; 34845d4f98a2SYan Zheng 34850257bb82SYan, Zheng err = __insert_orphan_inode(trans, root, objectid); 34865d4f98a2SYan Zheng BUG_ON(err); 34875d4f98a2SYan Zheng 34885d4f98a2SYan Zheng key.objectid = objectid; 34895d4f98a2SYan Zheng key.type = BTRFS_INODE_ITEM_KEY; 34905d4f98a2SYan Zheng key.offset = 0; 349173f73415SJosef Bacik inode = btrfs_iget(root->fs_info->sb, &key, root, NULL); 34925d4f98a2SYan Zheng BUG_ON(IS_ERR(inode) || is_bad_inode(inode)); 34935d4f98a2SYan Zheng BTRFS_I(inode)->index_cnt = group->key.objectid; 34945d4f98a2SYan Zheng 34955d4f98a2SYan Zheng err = btrfs_orphan_add(trans, inode); 34965d4f98a2SYan Zheng out: 34975d4f98a2SYan Zheng nr = trans->blocks_used; 34985d4f98a2SYan Zheng btrfs_end_transaction(trans, root); 34995d4f98a2SYan Zheng 35005d4f98a2SYan Zheng btrfs_btree_balance_dirty(root, nr); 35015d4f98a2SYan Zheng if (err) { 35025d4f98a2SYan Zheng if (inode) 35035d4f98a2SYan Zheng iput(inode); 35045d4f98a2SYan Zheng inode = ERR_PTR(err); 35055d4f98a2SYan Zheng } 35065d4f98a2SYan Zheng return inode; 35075d4f98a2SYan Zheng } 35085d4f98a2SYan Zheng 35095d4f98a2SYan Zheng /* 35105d4f98a2SYan Zheng * function to relocate all extents in a block group. 35115d4f98a2SYan Zheng */ 35125d4f98a2SYan Zheng int btrfs_relocate_block_group(struct btrfs_root *extent_root, u64 group_start) 35135d4f98a2SYan Zheng { 35145d4f98a2SYan Zheng struct btrfs_fs_info *fs_info = extent_root->fs_info; 35155d4f98a2SYan Zheng struct reloc_control *rc; 35165d4f98a2SYan Zheng int ret; 35175d4f98a2SYan Zheng int err = 0; 35185d4f98a2SYan Zheng 35195d4f98a2SYan Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 35205d4f98a2SYan Zheng if (!rc) 35215d4f98a2SYan Zheng return -ENOMEM; 35225d4f98a2SYan Zheng 35235d4f98a2SYan Zheng mapping_tree_init(&rc->reloc_root_tree); 35245d4f98a2SYan Zheng extent_io_tree_init(&rc->processed_blocks, NULL, GFP_NOFS); 35255d4f98a2SYan Zheng INIT_LIST_HEAD(&rc->reloc_roots); 35265d4f98a2SYan Zheng 35275d4f98a2SYan Zheng rc->block_group = btrfs_lookup_block_group(fs_info, group_start); 35285d4f98a2SYan Zheng BUG_ON(!rc->block_group); 35295d4f98a2SYan Zheng 35305d4f98a2SYan Zheng btrfs_init_workers(&rc->workers, "relocate", 353161d92c32SChris Mason fs_info->thread_pool_size, NULL); 35325d4f98a2SYan Zheng 35335d4f98a2SYan Zheng rc->extent_root = extent_root; 35345d4f98a2SYan Zheng btrfs_prepare_block_group_relocation(extent_root, rc->block_group); 35355d4f98a2SYan Zheng 35365d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, rc->block_group); 35375d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 35385d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 35395d4f98a2SYan Zheng rc->data_inode = NULL; 35405d4f98a2SYan Zheng goto out; 35415d4f98a2SYan Zheng } 35425d4f98a2SYan Zheng 35435d4f98a2SYan Zheng printk(KERN_INFO "btrfs: relocating block group %llu flags %llu\n", 35445d4f98a2SYan Zheng (unsigned long long)rc->block_group->key.objectid, 35455d4f98a2SYan Zheng (unsigned long long)rc->block_group->flags); 35465d4f98a2SYan Zheng 354724bbcf04SYan, Zheng btrfs_start_delalloc_inodes(fs_info->tree_root, 0); 354824bbcf04SYan, Zheng btrfs_wait_ordered_extents(fs_info->tree_root, 0, 0); 35495d4f98a2SYan Zheng 35505d4f98a2SYan Zheng while (1) { 35515d4f98a2SYan Zheng rc->extents_found = 0; 35525d4f98a2SYan Zheng rc->extents_skipped = 0; 35535d4f98a2SYan Zheng 355476dda93cSYan, Zheng mutex_lock(&fs_info->cleaner_mutex); 355576dda93cSYan, Zheng 355676dda93cSYan, Zheng btrfs_clean_old_snapshots(fs_info->tree_root); 35575d4f98a2SYan Zheng ret = relocate_block_group(rc); 355876dda93cSYan, Zheng 355976dda93cSYan, Zheng mutex_unlock(&fs_info->cleaner_mutex); 35605d4f98a2SYan Zheng if (ret < 0) { 35615d4f98a2SYan Zheng err = ret; 35625d4f98a2SYan Zheng break; 35635d4f98a2SYan Zheng } 35645d4f98a2SYan Zheng 35655d4f98a2SYan Zheng if (rc->extents_found == 0) 35665d4f98a2SYan Zheng break; 35675d4f98a2SYan Zheng 35685d4f98a2SYan Zheng printk(KERN_INFO "btrfs: found %llu extents\n", 35695d4f98a2SYan Zheng (unsigned long long)rc->extents_found); 35705d4f98a2SYan Zheng 35715d4f98a2SYan Zheng if (rc->stage == MOVE_DATA_EXTENTS && rc->found_file_extent) { 35725d4f98a2SYan Zheng btrfs_wait_ordered_range(rc->data_inode, 0, (u64)-1); 35735d4f98a2SYan Zheng invalidate_mapping_pages(rc->data_inode->i_mapping, 35745d4f98a2SYan Zheng 0, -1); 35755d4f98a2SYan Zheng rc->stage = UPDATE_DATA_PTRS; 35765d4f98a2SYan Zheng } else if (rc->stage == UPDATE_DATA_PTRS && 35775d4f98a2SYan Zheng rc->extents_skipped >= rc->extents_found) { 35785d4f98a2SYan Zheng iput(rc->data_inode); 35795d4f98a2SYan Zheng rc->data_inode = create_reloc_inode(fs_info, 35805d4f98a2SYan Zheng rc->block_group); 35815d4f98a2SYan Zheng if (IS_ERR(rc->data_inode)) { 35825d4f98a2SYan Zheng err = PTR_ERR(rc->data_inode); 35835d4f98a2SYan Zheng rc->data_inode = NULL; 35845d4f98a2SYan Zheng break; 35855d4f98a2SYan Zheng } 35865d4f98a2SYan Zheng rc->stage = MOVE_DATA_EXTENTS; 35875d4f98a2SYan Zheng rc->found_file_extent = 0; 35885d4f98a2SYan Zheng } 35895d4f98a2SYan Zheng } 35905d4f98a2SYan Zheng 35910257bb82SYan, Zheng filemap_write_and_wait_range(fs_info->btree_inode->i_mapping, 35925d4f98a2SYan Zheng rc->block_group->key.objectid, 35935d4f98a2SYan Zheng rc->block_group->key.objectid + 35945d4f98a2SYan Zheng rc->block_group->key.offset - 1); 35955d4f98a2SYan Zheng 35965d4f98a2SYan Zheng WARN_ON(rc->block_group->pinned > 0); 35975d4f98a2SYan Zheng WARN_ON(rc->block_group->reserved > 0); 35985d4f98a2SYan Zheng WARN_ON(btrfs_block_group_used(&rc->block_group->item) > 0); 35995d4f98a2SYan Zheng out: 36005d4f98a2SYan Zheng iput(rc->data_inode); 36015d4f98a2SYan Zheng btrfs_stop_workers(&rc->workers); 36025d4f98a2SYan Zheng btrfs_put_block_group(rc->block_group); 36035d4f98a2SYan Zheng kfree(rc); 36045d4f98a2SYan Zheng return err; 36055d4f98a2SYan Zheng } 36065d4f98a2SYan Zheng 360776dda93cSYan, Zheng static noinline_for_stack int mark_garbage_root(struct btrfs_root *root) 360876dda93cSYan, Zheng { 360976dda93cSYan, Zheng struct btrfs_trans_handle *trans; 361076dda93cSYan, Zheng int ret; 361176dda93cSYan, Zheng 361276dda93cSYan, Zheng trans = btrfs_start_transaction(root->fs_info->tree_root, 1); 361376dda93cSYan, Zheng 361476dda93cSYan, Zheng memset(&root->root_item.drop_progress, 0, 361576dda93cSYan, Zheng sizeof(root->root_item.drop_progress)); 361676dda93cSYan, Zheng root->root_item.drop_level = 0; 361776dda93cSYan, Zheng btrfs_set_root_refs(&root->root_item, 0); 361876dda93cSYan, Zheng ret = btrfs_update_root(trans, root->fs_info->tree_root, 361976dda93cSYan, Zheng &root->root_key, &root->root_item); 362076dda93cSYan, Zheng BUG_ON(ret); 362176dda93cSYan, Zheng 362276dda93cSYan, Zheng ret = btrfs_end_transaction(trans, root->fs_info->tree_root); 362376dda93cSYan, Zheng BUG_ON(ret); 362476dda93cSYan, Zheng return 0; 362576dda93cSYan, Zheng } 362676dda93cSYan, Zheng 36275d4f98a2SYan Zheng /* 36285d4f98a2SYan Zheng * recover relocation interrupted by system crash. 36295d4f98a2SYan Zheng * 36305d4f98a2SYan Zheng * this function resumes merging reloc trees with corresponding fs trees. 36315d4f98a2SYan Zheng * this is important for keeping the sharing of tree blocks 36325d4f98a2SYan Zheng */ 36335d4f98a2SYan Zheng int btrfs_recover_relocation(struct btrfs_root *root) 36345d4f98a2SYan Zheng { 36355d4f98a2SYan Zheng LIST_HEAD(reloc_roots); 36365d4f98a2SYan Zheng struct btrfs_key key; 36375d4f98a2SYan Zheng struct btrfs_root *fs_root; 36385d4f98a2SYan Zheng struct btrfs_root *reloc_root; 36395d4f98a2SYan Zheng struct btrfs_path *path; 36405d4f98a2SYan Zheng struct extent_buffer *leaf; 36415d4f98a2SYan Zheng struct reloc_control *rc = NULL; 36425d4f98a2SYan Zheng struct btrfs_trans_handle *trans; 36435d4f98a2SYan Zheng int ret; 36445d4f98a2SYan Zheng int err = 0; 36455d4f98a2SYan Zheng 36465d4f98a2SYan Zheng path = btrfs_alloc_path(); 36475d4f98a2SYan Zheng if (!path) 36485d4f98a2SYan Zheng return -ENOMEM; 36495d4f98a2SYan Zheng 36505d4f98a2SYan Zheng key.objectid = BTRFS_TREE_RELOC_OBJECTID; 36515d4f98a2SYan Zheng key.type = BTRFS_ROOT_ITEM_KEY; 36525d4f98a2SYan Zheng key.offset = (u64)-1; 36535d4f98a2SYan Zheng 36545d4f98a2SYan Zheng while (1) { 36555d4f98a2SYan Zheng ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, 36565d4f98a2SYan Zheng path, 0, 0); 36575d4f98a2SYan Zheng if (ret < 0) { 36585d4f98a2SYan Zheng err = ret; 36595d4f98a2SYan Zheng goto out; 36605d4f98a2SYan Zheng } 36615d4f98a2SYan Zheng if (ret > 0) { 36625d4f98a2SYan Zheng if (path->slots[0] == 0) 36635d4f98a2SYan Zheng break; 36645d4f98a2SYan Zheng path->slots[0]--; 36655d4f98a2SYan Zheng } 36665d4f98a2SYan Zheng leaf = path->nodes[0]; 36675d4f98a2SYan Zheng btrfs_item_key_to_cpu(leaf, &key, path->slots[0]); 36685d4f98a2SYan Zheng btrfs_release_path(root->fs_info->tree_root, path); 36695d4f98a2SYan Zheng 36705d4f98a2SYan Zheng if (key.objectid != BTRFS_TREE_RELOC_OBJECTID || 36715d4f98a2SYan Zheng key.type != BTRFS_ROOT_ITEM_KEY) 36725d4f98a2SYan Zheng break; 36735d4f98a2SYan Zheng 36745d4f98a2SYan Zheng reloc_root = btrfs_read_fs_root_no_radix(root, &key); 36755d4f98a2SYan Zheng if (IS_ERR(reloc_root)) { 36765d4f98a2SYan Zheng err = PTR_ERR(reloc_root); 36775d4f98a2SYan Zheng goto out; 36785d4f98a2SYan Zheng } 36795d4f98a2SYan Zheng 36805d4f98a2SYan Zheng list_add(&reloc_root->root_list, &reloc_roots); 36815d4f98a2SYan Zheng 36825d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) > 0) { 36835d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 36845d4f98a2SYan Zheng reloc_root->root_key.offset); 36855d4f98a2SYan Zheng if (IS_ERR(fs_root)) { 368676dda93cSYan, Zheng ret = PTR_ERR(fs_root); 368776dda93cSYan, Zheng if (ret != -ENOENT) { 368876dda93cSYan, Zheng err = ret; 36895d4f98a2SYan Zheng goto out; 36905d4f98a2SYan Zheng } 369176dda93cSYan, Zheng mark_garbage_root(reloc_root); 369276dda93cSYan, Zheng } 36935d4f98a2SYan Zheng } 36945d4f98a2SYan Zheng 36955d4f98a2SYan Zheng if (key.offset == 0) 36965d4f98a2SYan Zheng break; 36975d4f98a2SYan Zheng 36985d4f98a2SYan Zheng key.offset--; 36995d4f98a2SYan Zheng } 37005d4f98a2SYan Zheng btrfs_release_path(root->fs_info->tree_root, path); 37015d4f98a2SYan Zheng 37025d4f98a2SYan Zheng if (list_empty(&reloc_roots)) 37035d4f98a2SYan Zheng goto out; 37045d4f98a2SYan Zheng 37055d4f98a2SYan Zheng rc = kzalloc(sizeof(*rc), GFP_NOFS); 37065d4f98a2SYan Zheng if (!rc) { 37075d4f98a2SYan Zheng err = -ENOMEM; 37085d4f98a2SYan Zheng goto out; 37095d4f98a2SYan Zheng } 37105d4f98a2SYan Zheng 37115d4f98a2SYan Zheng mapping_tree_init(&rc->reloc_root_tree); 37125d4f98a2SYan Zheng INIT_LIST_HEAD(&rc->reloc_roots); 37135d4f98a2SYan Zheng btrfs_init_workers(&rc->workers, "relocate", 371461d92c32SChris Mason root->fs_info->thread_pool_size, NULL); 37155d4f98a2SYan Zheng rc->extent_root = root->fs_info->extent_root; 37165d4f98a2SYan Zheng 37175d4f98a2SYan Zheng set_reloc_control(rc); 37185d4f98a2SYan Zheng 37195d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 37205d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 37215d4f98a2SYan Zheng struct btrfs_root, root_list); 37225d4f98a2SYan Zheng list_del(&reloc_root->root_list); 37235d4f98a2SYan Zheng 37245d4f98a2SYan Zheng if (btrfs_root_refs(&reloc_root->root_item) == 0) { 37255d4f98a2SYan Zheng list_add_tail(&reloc_root->root_list, 37265d4f98a2SYan Zheng &rc->reloc_roots); 37275d4f98a2SYan Zheng continue; 37285d4f98a2SYan Zheng } 37295d4f98a2SYan Zheng 37305d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 37315d4f98a2SYan Zheng reloc_root->root_key.offset); 37325d4f98a2SYan Zheng BUG_ON(IS_ERR(fs_root)); 37335d4f98a2SYan Zheng 37345d4f98a2SYan Zheng __add_reloc_root(reloc_root); 37355d4f98a2SYan Zheng fs_root->reloc_root = reloc_root; 37365d4f98a2SYan Zheng } 37375d4f98a2SYan Zheng 37385d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 37395d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 37405d4f98a2SYan Zheng 37415d4f98a2SYan Zheng merge_reloc_roots(rc); 37425d4f98a2SYan Zheng 37435d4f98a2SYan Zheng unset_reloc_control(rc); 37445d4f98a2SYan Zheng 37455d4f98a2SYan Zheng trans = btrfs_start_transaction(rc->extent_root, 1); 37465d4f98a2SYan Zheng btrfs_commit_transaction(trans, rc->extent_root); 37475d4f98a2SYan Zheng out: 37485d4f98a2SYan Zheng if (rc) { 37495d4f98a2SYan Zheng btrfs_stop_workers(&rc->workers); 37505d4f98a2SYan Zheng kfree(rc); 37515d4f98a2SYan Zheng } 37525d4f98a2SYan Zheng while (!list_empty(&reloc_roots)) { 37535d4f98a2SYan Zheng reloc_root = list_entry(reloc_roots.next, 37545d4f98a2SYan Zheng struct btrfs_root, root_list); 37555d4f98a2SYan Zheng list_del(&reloc_root->root_list); 37565d4f98a2SYan Zheng free_extent_buffer(reloc_root->node); 37575d4f98a2SYan Zheng free_extent_buffer(reloc_root->commit_root); 37585d4f98a2SYan Zheng kfree(reloc_root); 37595d4f98a2SYan Zheng } 37605d4f98a2SYan Zheng btrfs_free_path(path); 37615d4f98a2SYan Zheng 37625d4f98a2SYan Zheng if (err == 0) { 37635d4f98a2SYan Zheng /* cleanup orphan inode in data relocation tree */ 37645d4f98a2SYan Zheng fs_root = read_fs_root(root->fs_info, 37655d4f98a2SYan Zheng BTRFS_DATA_RELOC_TREE_OBJECTID); 37665d4f98a2SYan Zheng if (IS_ERR(fs_root)) 37675d4f98a2SYan Zheng err = PTR_ERR(fs_root); 3768d7ce5843SMiao Xie else 3769c71bf099SYan, Zheng btrfs_orphan_cleanup(fs_root); 37705d4f98a2SYan Zheng } 37715d4f98a2SYan Zheng return err; 37725d4f98a2SYan Zheng } 37735d4f98a2SYan Zheng 37745d4f98a2SYan Zheng /* 37755d4f98a2SYan Zheng * helper to add ordered checksum for data relocation. 37765d4f98a2SYan Zheng * 37775d4f98a2SYan Zheng * cloning checksum properly handles the nodatasum extents. 37785d4f98a2SYan Zheng * it also saves CPU time to re-calculate the checksum. 37795d4f98a2SYan Zheng */ 37805d4f98a2SYan Zheng int btrfs_reloc_clone_csums(struct inode *inode, u64 file_pos, u64 len) 37815d4f98a2SYan Zheng { 37825d4f98a2SYan Zheng struct btrfs_ordered_sum *sums; 37835d4f98a2SYan Zheng struct btrfs_sector_sum *sector_sum; 37845d4f98a2SYan Zheng struct btrfs_ordered_extent *ordered; 37855d4f98a2SYan Zheng struct btrfs_root *root = BTRFS_I(inode)->root; 37865d4f98a2SYan Zheng size_t offset; 37875d4f98a2SYan Zheng int ret; 37885d4f98a2SYan Zheng u64 disk_bytenr; 37895d4f98a2SYan Zheng LIST_HEAD(list); 37905d4f98a2SYan Zheng 37915d4f98a2SYan Zheng ordered = btrfs_lookup_ordered_extent(inode, file_pos); 37925d4f98a2SYan Zheng BUG_ON(ordered->file_offset != file_pos || ordered->len != len); 37935d4f98a2SYan Zheng 37945d4f98a2SYan Zheng disk_bytenr = file_pos + BTRFS_I(inode)->index_cnt; 37955d4f98a2SYan Zheng ret = btrfs_lookup_csums_range(root->fs_info->csum_root, disk_bytenr, 37965d4f98a2SYan Zheng disk_bytenr + len - 1, &list); 37975d4f98a2SYan Zheng 37985d4f98a2SYan Zheng while (!list_empty(&list)) { 37995d4f98a2SYan Zheng sums = list_entry(list.next, struct btrfs_ordered_sum, list); 38005d4f98a2SYan Zheng list_del_init(&sums->list); 38015d4f98a2SYan Zheng 38025d4f98a2SYan Zheng sector_sum = sums->sums; 38035d4f98a2SYan Zheng sums->bytenr = ordered->start; 38045d4f98a2SYan Zheng 38055d4f98a2SYan Zheng offset = 0; 38065d4f98a2SYan Zheng while (offset < sums->len) { 38075d4f98a2SYan Zheng sector_sum->bytenr += ordered->start - disk_bytenr; 38085d4f98a2SYan Zheng sector_sum++; 38095d4f98a2SYan Zheng offset += root->sectorsize; 38105d4f98a2SYan Zheng } 38115d4f98a2SYan Zheng 38125d4f98a2SYan Zheng btrfs_add_ordered_sum(inode, ordered, sums); 38135d4f98a2SYan Zheng } 38145d4f98a2SYan Zheng btrfs_put_ordered_extent(ordered); 38155d4f98a2SYan Zheng return 0; 38165d4f98a2SYan Zheng } 3817