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 (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
31 		goto out;
32 
33 	path = btrfs_alloc_path();
34 	if (!path)
35 		return -ENOMEM;
36 
37 	level = btrfs_header_level(root->node);
38 
39 	if (level == 0)
40 		goto out;
41 
42 	if (root->defrag_progress.objectid == 0) {
43 		struct extent_buffer *root_node;
44 		u32 nritems;
45 
46 		root_node = btrfs_lock_root_node(root);
47 		nritems = btrfs_header_nritems(root_node);
48 		root->defrag_max.objectid = 0;
49 		/* from above we know this is not a leaf */
50 		btrfs_node_key_to_cpu(root_node, &root->defrag_max,
51 				      nritems - 1);
52 		btrfs_tree_unlock(root_node);
53 		free_extent_buffer(root_node);
54 		memset(&key, 0, sizeof(key));
55 	} else {
56 		memcpy(&key, &root->defrag_progress, sizeof(key));
57 	}
58 
59 	path->keep_locks = 1;
60 
61 	ret = btrfs_search_forward(root, &key, path, BTRFS_OLDEST_GENERATION);
62 	if (ret < 0)
63 		goto out;
64 	if (ret > 0) {
65 		ret = 0;
66 		goto out;
67 	}
68 	btrfs_release_path(path);
69 	/*
70 	 * We don't need a lock on a leaf. btrfs_realloc_node() will lock all
71 	 * leafs from path->nodes[1], so set lowest_level to 1 to avoid later
72 	 * a deadlock (attempting to write lock an already write locked leaf).
73 	 */
74 	path->lowest_level = 1;
75 	wret = btrfs_search_slot(trans, root, &key, path, 0, 1);
76 
77 	if (wret < 0) {
78 		ret = wret;
79 		goto out;
80 	}
81 	if (!path->nodes[1]) {
82 		ret = 0;
83 		goto out;
84 	}
85 	/*
86 	 * The node at level 1 must always be locked when our path has
87 	 * keep_locks set and lowest_level is 1, regardless of the value of
88 	 * path->slots[1].
89 	 */
90 	BUG_ON(path->locks[1] == 0);
91 	ret = btrfs_realloc_node(trans, root,
92 				 path->nodes[1], 0,
93 				 &last_ret,
94 				 &root->defrag_progress);
95 	if (ret) {
96 		WARN_ON(ret == -EAGAIN);
97 		goto out;
98 	}
99 	/*
100 	 * Now that we reallocated the node we can find the next key. Note that
101 	 * btrfs_find_next_key() can release our path and do another search
102 	 * without COWing, this is because even with path->keep_locks = 1,
103 	 * btrfs_search_slot() / ctree.c:unlock_up() does not keeps a lock on a
104 	 * node when path->slots[node_level - 1] does not point to the last
105 	 * item or a slot beyond the last item (ctree.c:unlock_up()). Therefore
106 	 * we search for the next key after reallocating our node.
107 	 */
108 	path->slots[1] = btrfs_header_nritems(path->nodes[1]);
109 	next_key_ret = btrfs_find_next_key(root, path, &key, 1,
110 					   BTRFS_OLDEST_GENERATION);
111 	if (next_key_ret == 0) {
112 		memcpy(&root->defrag_progress, &key, sizeof(key));
113 		ret = -EAGAIN;
114 	}
115 out:
116 	btrfs_free_path(path);
117 	if (ret == -EAGAIN) {
118 		if (root->defrag_max.objectid > root->defrag_progress.objectid)
119 			goto done;
120 		if (root->defrag_max.type > root->defrag_progress.type)
121 			goto done;
122 		if (root->defrag_max.offset > root->defrag_progress.offset)
123 			goto done;
124 		ret = 0;
125 	}
126 done:
127 	if (ret != -EAGAIN)
128 		memset(&root->defrag_progress, 0,
129 		       sizeof(root->defrag_progress));
130 
131 	return ret;
132 }
133