1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6 #include <linux/sched.h>
7 #include "ctree.h"
8 #include "disk-io.h"
9 #include "print-tree.h"
10 #include "transaction.h"
11 #include "locking.h"
12
13 /*
14 * Defrag all the leaves in a given btree.
15 * Read all the leaves and try to get key order to
16 * better reflect disk order
17 */
18
btrfs_defrag_leaves(struct btrfs_trans_handle * trans,struct btrfs_root * root)19 int btrfs_defrag_leaves(struct btrfs_trans_handle *trans,
20 struct btrfs_root *root)
21 {
22 struct btrfs_path *path = NULL;
23 struct btrfs_key key;
24 int ret = 0;
25 int wret;
26 int level;
27 int next_key_ret = 0;
28 u64 last_ret = 0;
29
30 if (root->fs_info->extent_root == root) {
31 /*
32 * there's recursion here right now in the tree locking,
33 * we can't defrag the extent root without deadlock
34 */
35 goto out;
36 }
37
38 if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
39 goto out;
40
41 path = btrfs_alloc_path();
42 if (!path) {
43 ret = -ENOMEM;
44 goto out;
45 }
46
47 level = btrfs_header_level(root->node);
48
49 if (level == 0)
50 goto out;
51
52 if (root->defrag_progress.objectid == 0) {
53 struct extent_buffer *root_node;
54 u32 nritems;
55
56 root_node = btrfs_lock_root_node(root);
57 nritems = btrfs_header_nritems(root_node);
58 root->defrag_max.objectid = 0;
59 /* from above we know this is not a leaf */
60 btrfs_node_key_to_cpu(root_node, &root->defrag_max,
61 nritems - 1);
62 btrfs_tree_unlock(root_node);
63 free_extent_buffer(root_node);
64 memset(&key, 0, sizeof(key));
65 } else {
66 memcpy(&key, &root->defrag_progress, sizeof(key));
67 }
68
69 path->keep_locks = 1;
70
71 ret = btrfs_search_forward(root, &key, path, BTRFS_OLDEST_GENERATION);
72 if (ret < 0)
73 goto out;
74 if (ret > 0) {
75 ret = 0;
76 goto out;
77 }
78 btrfs_release_path(path);
79 /*
80 * We don't need a lock on a leaf. btrfs_realloc_node() will lock all
81 * leafs from path->nodes[1], so set lowest_level to 1 to avoid later
82 * a deadlock (attempting to write lock an already write locked leaf).
83 */
84 path->lowest_level = 1;
85 wret = btrfs_search_slot(trans, root, &key, path, 0, 1);
86
87 if (wret < 0) {
88 ret = wret;
89 goto out;
90 }
91 if (!path->nodes[1]) {
92 ret = 0;
93 goto out;
94 }
95 /*
96 * The node at level 1 must always be locked when our path has
97 * keep_locks set and lowest_level is 1, regardless of the value of
98 * path->slots[1].
99 */
100 BUG_ON(path->locks[1] == 0);
101 ret = btrfs_realloc_node(trans, root,
102 path->nodes[1], 0,
103 &last_ret,
104 &root->defrag_progress);
105 if (ret) {
106 WARN_ON(ret == -EAGAIN);
107 goto out;
108 }
109 /*
110 * Now that we reallocated the node we can find the next key. Note that
111 * btrfs_find_next_key() can release our path and do another search
112 * without COWing, this is because even with path->keep_locks = 1,
113 * btrfs_search_slot() / ctree.c:unlock_up() does not keeps a lock on a
114 * node when path->slots[node_level - 1] does not point to the last
115 * item or a slot beyond the last item (ctree.c:unlock_up()). Therefore
116 * we search for the next key after reallocating our node.
117 */
118 path->slots[1] = btrfs_header_nritems(path->nodes[1]);
119 next_key_ret = btrfs_find_next_key(root, path, &key, 1,
120 BTRFS_OLDEST_GENERATION);
121 if (next_key_ret == 0) {
122 memcpy(&root->defrag_progress, &key, sizeof(key));
123 ret = -EAGAIN;
124 }
125 out:
126 btrfs_free_path(path);
127 if (ret == -EAGAIN) {
128 if (root->defrag_max.objectid > root->defrag_progress.objectid)
129 goto done;
130 if (root->defrag_max.type > root->defrag_progress.type)
131 goto done;
132 if (root->defrag_max.offset > root->defrag_progress.offset)
133 goto done;
134 ret = 0;
135 }
136 done:
137 if (ret != -EAGAIN)
138 memset(&root->defrag_progress, 0,
139 sizeof(root->defrag_progress));
140
141 return ret;
142 }
143