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
 19int 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		return -ENOMEM;
 44
 45	level = btrfs_header_level(root->node);
 46
 47	if (level == 0)
 48		goto out;
 49
 50	if (root->defrag_progress.objectid == 0) {
 51		struct extent_buffer *root_node;
 52		u32 nritems;
 53
 54		root_node = btrfs_lock_root_node(root);
 55		btrfs_set_lock_blocking_write(root_node);
 56		nritems = btrfs_header_nritems(root_node);
 57		root->defrag_max.objectid = 0;
 58		/* from above we know this is not a leaf */
 59		btrfs_node_key_to_cpu(root_node, &root->defrag_max,
 60				      nritems - 1);
 61		btrfs_tree_unlock(root_node);
 62		free_extent_buffer(root_node);
 63		memset(&key, 0, sizeof(key));
 64	} else {
 65		memcpy(&key, &root->defrag_progress, sizeof(key));
 66	}
 67
 68	path->keep_locks = 1;
 69
 70	ret = btrfs_search_forward(root, &key, path, BTRFS_OLDEST_GENERATION);
 71	if (ret < 0)
 72		goto out;
 73	if (ret > 0) {
 74		ret = 0;
 75		goto out;
 76	}
 77	btrfs_release_path(path);
 78	/*
 79	 * We don't need a lock on a leaf. btrfs_realloc_node() will lock all
 80	 * leafs from path->nodes[1], so set lowest_level to 1 to avoid later
 81	 * a deadlock (attempting to write lock an already write locked leaf).
 82	 */
 83	path->lowest_level = 1;
 84	wret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 85
 86	if (wret < 0) {
 87		ret = wret;
 88		goto out;
 89	}
 90	if (!path->nodes[1]) {
 91		ret = 0;
 92		goto out;
 93	}
 94	/*
 95	 * The node at level 1 must always be locked when our path has
 96	 * keep_locks set and lowest_level is 1, regardless of the value of
 97	 * path->slots[1].
 98	 */
 99	BUG_ON(path->locks[1] == 0);
100	ret = btrfs_realloc_node(trans, root,
101				 path->nodes[1], 0,
102				 &last_ret,
103				 &root->defrag_progress);
104	if (ret) {
105		WARN_ON(ret == -EAGAIN);
106		goto out;
107	}
108	/*
109	 * Now that we reallocated the node we can find the next key. Note that
110	 * btrfs_find_next_key() can release our path and do another search
111	 * without COWing, this is because even with path->keep_locks = 1,
112	 * btrfs_search_slot() / ctree.c:unlock_up() does not keeps a lock on a
113	 * node when path->slots[node_level - 1] does not point to the last
114	 * item or a slot beyond the last item (ctree.c:unlock_up()). Therefore
115	 * we search for the next key after reallocating our node.
116	 */
117	path->slots[1] = btrfs_header_nritems(path->nodes[1]);
118	next_key_ret = btrfs_find_next_key(root, path, &key, 1,
119					   BTRFS_OLDEST_GENERATION);
120	if (next_key_ret == 0) {
121		memcpy(&root->defrag_progress, &key, sizeof(key));
122		ret = -EAGAIN;
123	}
124out:
125	btrfs_free_path(path);
126	if (ret == -EAGAIN) {
127		if (root->defrag_max.objectid > root->defrag_progress.objectid)
128			goto done;
129		if (root->defrag_max.type > root->defrag_progress.type)
130			goto done;
131		if (root->defrag_max.offset > root->defrag_progress.offset)
132			goto done;
133		ret = 0;
134	}
135done:
136	if (ret != -EAGAIN)
137		memset(&root->defrag_progress, 0,
138		       sizeof(root->defrag_progress));
139
140	return ret;
141}