xref: /linux/fs/btrfs/free-space-tree.c (revision 79bdd8846317f3dea26c53d75700045f62265557)

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2015 Facebook.  All rights reserved.
 */

#include <linux/kernel.h>
#include <linux/sched/mm.h>
#include "messages.h"
#include "ctree.h"
#include "disk-io.h"
#include "locking.h"
#include "free-space-tree.h"
#include "transaction.h"
#include "block-group.h"
#include "fs.h"
#include "accessors.h"
#include "extent-tree.h"
#include "root-tree.h"

static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
					struct btrfs_block_group *block_group,
					struct btrfs_path *path);

struct btrfs_root *btrfs_free_space_root(struct btrfs_block_group *block_group)
{
	struct btrfs_key key = {
		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
		.type = BTRFS_ROOT_ITEM_KEY,
		.offset = 0,
	};

	if (btrfs_fs_incompat(block_group->fs_info, EXTENT_TREE_V2))
		key.offset = block_group->global_root_id;
	return btrfs_global_root(block_group->fs_info, &key);
}

void btrfs_set_free_space_tree_thresholds(struct btrfs_block_group *cache)
{
	u32 bitmap_range;
	size_t bitmap_size;
	u64 num_bitmaps, total_bitmap_size;

	if (WARN_ON(cache->length == 0))
		btrfs_warn(cache->fs_info, "block group %llu length is zero",
			   cache->start);

	/*
	 * We convert to bitmaps when the disk space required for using extents
	 * exceeds that required for using bitmaps.
	 */
	bitmap_range = cache->fs_info->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
	num_bitmaps = div_u64(cache->length + bitmap_range - 1, bitmap_range);
	bitmap_size = sizeof(struct btrfs_item) + BTRFS_FREE_SPACE_BITMAP_SIZE;
	total_bitmap_size = num_bitmaps * bitmap_size;
	cache->bitmap_high_thresh = div_u64(total_bitmap_size,
					    sizeof(struct btrfs_item));

	/*
	 * We allow for a small buffer between the high threshold and low
	 * threshold to avoid thrashing back and forth between the two formats.
	 */
	if (cache->bitmap_high_thresh > 100)
		cache->bitmap_low_thresh = cache->bitmap_high_thresh - 100;
	else
		cache->bitmap_low_thresh = 0;
}

static int add_new_free_space_info(struct btrfs_trans_handle *trans,
				   struct btrfs_block_group *block_group,
				   struct btrfs_path *path)
{
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	struct btrfs_free_space_info *info;
	struct btrfs_key key;
	struct extent_buffer *leaf;
	int ret;

	key.objectid = block_group->start;
	key.type = BTRFS_FREE_SPACE_INFO_KEY;
	key.offset = block_group->length;

	ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*info));
	if (ret)
		return ret;

	leaf = path->nodes[0];
	info = btrfs_item_ptr(leaf, path->slots[0],
			      struct btrfs_free_space_info);
	btrfs_set_free_space_extent_count(leaf, info, 0);
	btrfs_set_free_space_flags(leaf, info, 0);
	btrfs_release_path(path);
	return 0;
}

struct btrfs_free_space_info *btrfs_search_free_space_info(
		struct btrfs_trans_handle *trans,
		struct btrfs_block_group *block_group,
		struct btrfs_path *path, int cow)
{
	struct btrfs_fs_info *fs_info = block_group->fs_info;
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	struct btrfs_key key;
	int ret;

	key.objectid = block_group->start;
	key.type = BTRFS_FREE_SPACE_INFO_KEY;
	key.offset = block_group->length;

	ret = btrfs_search_slot(trans, root, &key, path, 0, cow);
	if (ret < 0)
		return ERR_PTR(ret);
	if (unlikely(ret != 0)) {
		btrfs_warn(fs_info, "missing free space info for %llu",
			   block_group->start);
		DEBUG_WARN();
		return ERR_PTR(-ENOENT);
	}

	return btrfs_item_ptr(path->nodes[0], path->slots[0],
			      struct btrfs_free_space_info);
}

/*
 * btrfs_search_slot() but we're looking for the greatest key less than the
 * passed key.
 */
static int btrfs_search_prev_slot(struct btrfs_trans_handle *trans,
				  struct btrfs_root *root,
				  struct btrfs_key *key, struct btrfs_path *p,
				  int ins_len, int cow)
{
	int ret;

	ret = btrfs_search_slot(trans, root, key, p, ins_len, cow);
	if (ret < 0)
		return ret;

	if (unlikely(ret == 0)) {
		DEBUG_WARN();
		return -EIO;
	}

	if (unlikely(p->slots[0] == 0)) {
		DEBUG_WARN("no previous slot found");
		return -EIO;
	}
	p->slots[0]--;

	return 0;
}

static inline u32 free_space_bitmap_size(const struct btrfs_fs_info *fs_info,
					 u64 size)
{
	return DIV_ROUND_UP(size >> fs_info->sectorsize_bits, BITS_PER_BYTE);
}

static unsigned long *alloc_bitmap(u32 bitmap_size)
{
	unsigned long *ret;
	unsigned int nofs_flag;
	u32 bitmap_rounded_size = round_up(bitmap_size, sizeof(unsigned long));

	/*
	 * GFP_NOFS doesn't work with kvmalloc(), but we really can't recurse
	 * into the filesystem here. All callers hold a transaction handle
	 * open, so if a GFP_KERNEL allocation recurses into the filesystem
	 * and triggers a transaction commit, we would deadlock.
	 */
	nofs_flag = memalloc_nofs_save();
	ret = kvzalloc(bitmap_rounded_size, GFP_KERNEL);
	memalloc_nofs_restore(nofs_flag);
	return ret;
}

static void le_bitmap_set(unsigned long *map, unsigned int start, int len)
{
	u8 *p = ((u8 *)map) + BIT_BYTE(start);
	const unsigned int size = start + len;
	int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
	u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);

	while (len - bits_to_set >= 0) {
		*p |= mask_to_set;
		len -= bits_to_set;
		bits_to_set = BITS_PER_BYTE;
		mask_to_set = ~0;
		p++;
	}
	if (len) {
		mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
		*p |= mask_to_set;
	}
}

EXPORT_FOR_TESTS
int btrfs_convert_free_space_to_bitmaps(struct btrfs_trans_handle *trans,
					struct btrfs_block_group *block_group,
					struct btrfs_path *path)
{
	struct btrfs_fs_info *fs_info = trans->fs_info;
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	struct btrfs_free_space_info *info;
	struct btrfs_key key, found_key;
	struct extent_buffer *leaf;
	unsigned long *bitmap;
	char *bitmap_cursor;
	u64 start, end;
	u64 bitmap_range, i;
	u32 bitmap_size, flags, expected_extent_count;
	u32 extent_count = 0;
	bool done = false;
	int nr;
	int ret;

	bitmap_size = free_space_bitmap_size(fs_info, block_group->length);
	bitmap = alloc_bitmap(bitmap_size);
	if (unlikely(!bitmap))
		return 0;

	start = block_group->start;
	end = btrfs_block_group_end(block_group);

	key.objectid = end - 1;
	key.type = (u8)-1;
	key.offset = (u64)-1;

	while (!done) {
		ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			goto out;
		}

		leaf = path->nodes[0];
		nr = 0;
		path->slots[0]++;
		while (path->slots[0] > 0) {
			btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);

			if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
				ASSERT(found_key.objectid == block_group->start);
				ASSERT(found_key.offset == block_group->length);
				done = true;
				break;
			} else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY) {
				u64 first, last;

				ASSERT(found_key.objectid >= start);
				ASSERT(found_key.objectid < end);
				ASSERT(found_key.objectid + found_key.offset <= end);

				first = div_u64(found_key.objectid - start,
						fs_info->sectorsize);
				last = div_u64(found_key.objectid + found_key.offset - start,
					       fs_info->sectorsize);
				le_bitmap_set(bitmap, first, last - first);

				extent_count++;
				nr++;
				path->slots[0]--;
			} else {
				btrfs_err(fs_info, "unexpected free space tree key type %u",
					  found_key.type);
				ret = -EUCLEAN;
				btrfs_abort_transaction(trans, ret);
				goto out;
			}
		}

		ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			goto out;
		}
		btrfs_release_path(path);
	}

	info = btrfs_search_free_space_info(trans, block_group, path, 1);
	if (IS_ERR(info)) {
		ret = PTR_ERR(info);
		btrfs_abort_transaction(trans, ret);
		goto out;
	}
	leaf = path->nodes[0];
	flags = btrfs_free_space_flags(leaf, info);
	flags |= BTRFS_FREE_SPACE_USING_BITMAPS;
	block_group->using_free_space_bitmaps = true;
	block_group->using_free_space_bitmaps_cached = true;
	btrfs_set_free_space_flags(leaf, info, flags);
	expected_extent_count = btrfs_free_space_extent_count(leaf, info);
	btrfs_release_path(path);

	if (unlikely(extent_count != expected_extent_count)) {
		btrfs_err(fs_info,
			  "incorrect extent count for %llu; counted %u, expected %u",
			  block_group->start, extent_count,
			  expected_extent_count);
		ret = -EIO;
		btrfs_abort_transaction(trans, ret);
		goto out;
	}

	bitmap_cursor = (char *)bitmap;
	bitmap_range = fs_info->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
	i = start;
	while (i < end) {
		unsigned long ptr;
		u64 extent_size;
		u32 data_size;

		extent_size = min(end - i, bitmap_range);
		data_size = free_space_bitmap_size(fs_info, extent_size);

		key.objectid = i;
		key.type = BTRFS_FREE_SPACE_BITMAP_KEY;
		key.offset = extent_size;

		ret = btrfs_insert_empty_item(trans, root, path, &key,
					      data_size);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			goto out;
		}

		leaf = path->nodes[0];
		ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
		write_extent_buffer(leaf, bitmap_cursor, ptr,
				    data_size);
		btrfs_release_path(path);

		i += extent_size;
		bitmap_cursor += data_size;
	}

	ret = 0;
out:
	kvfree(bitmap);
	return ret;
}

EXPORT_FOR_TESTS
int btrfs_convert_free_space_to_extents(struct btrfs_trans_handle *trans,
					struct btrfs_block_group *block_group,
					struct btrfs_path *path)
{
	struct btrfs_fs_info *fs_info = trans->fs_info;
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	struct btrfs_free_space_info *info;
	struct btrfs_key key, found_key;
	struct extent_buffer *leaf;
	unsigned long *bitmap;
	u64 start, end;
	u32 bitmap_size, flags, expected_extent_count;
	unsigned long nrbits, start_bit, end_bit;
	u32 extent_count = 0;
	bool done = false;
	int nr;
	int ret;

	bitmap_size = free_space_bitmap_size(fs_info, block_group->length);
	bitmap = alloc_bitmap(bitmap_size);
	if (unlikely(!bitmap))
		return 0;

	start = block_group->start;
	end = btrfs_block_group_end(block_group);

	key.objectid = end - 1;
	key.type = (u8)-1;
	key.offset = (u64)-1;

	while (!done) {
		ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			goto out;
		}

		leaf = path->nodes[0];
		nr = 0;
		path->slots[0]++;
		while (path->slots[0] > 0) {
			btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);

			if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
				ASSERT(found_key.objectid == block_group->start);
				ASSERT(found_key.offset == block_group->length);
				done = true;
				break;
			} else if (found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
				unsigned long ptr;
				char *bitmap_cursor;
				u32 bitmap_pos, data_size;

				ASSERT(found_key.objectid >= start);
				ASSERT(found_key.objectid < end);
				ASSERT(found_key.objectid + found_key.offset <= end);

				bitmap_pos = div_u64(found_key.objectid - start,
						     fs_info->sectorsize *
						     BITS_PER_BYTE);
				bitmap_cursor = ((char *)bitmap) + bitmap_pos;
				data_size = free_space_bitmap_size(fs_info,
								found_key.offset);

				path->slots[0]--;
				ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
				read_extent_buffer(leaf, bitmap_cursor, ptr,
						   data_size);

				nr++;
			} else {
				btrfs_err(fs_info, "unexpected free space tree key type %u",
					  found_key.type);
				ret = -EUCLEAN;
				btrfs_abort_transaction(trans, ret);
				goto out;
			}
		}

		ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			goto out;
		}
		btrfs_release_path(path);
	}

	info = btrfs_search_free_space_info(trans, block_group, path, 1);
	if (IS_ERR(info)) {
		ret = PTR_ERR(info);
		btrfs_abort_transaction(trans, ret);
		goto out;
	}
	leaf = path->nodes[0];
	flags = btrfs_free_space_flags(leaf, info);
	flags &= ~BTRFS_FREE_SPACE_USING_BITMAPS;
	block_group->using_free_space_bitmaps = false;
	block_group->using_free_space_bitmaps_cached = true;
	btrfs_set_free_space_flags(leaf, info, flags);
	expected_extent_count = btrfs_free_space_extent_count(leaf, info);
	btrfs_release_path(path);

	nrbits = block_group->length >> fs_info->sectorsize_bits;
	start_bit = find_next_bit_le(bitmap, nrbits, 0);

	while (start_bit < nrbits) {
		end_bit = find_next_zero_bit_le(bitmap, nrbits, start_bit);
		ASSERT(start_bit < end_bit);

		key.objectid = start + start_bit * fs_info->sectorsize;
		key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
		key.offset = (end_bit - start_bit) * fs_info->sectorsize;

		ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			goto out;
		}
		btrfs_release_path(path);

		extent_count++;

		start_bit = find_next_bit_le(bitmap, nrbits, end_bit);
	}

	if (unlikely(extent_count != expected_extent_count)) {
		btrfs_err(fs_info,
			  "incorrect extent count for %llu; counted %u, expected %u",
			  block_group->start, extent_count,
			  expected_extent_count);
		ret = -EIO;
		btrfs_abort_transaction(trans, ret);
		goto out;
	}

	ret = 0;
out:
	kvfree(bitmap);
	return ret;
}

static int update_free_space_extent_count(struct btrfs_trans_handle *trans,
					  struct btrfs_block_group *block_group,
					  struct btrfs_path *path,
					  int new_extents)
{
	struct btrfs_free_space_info *info;
	u32 flags;
	u32 extent_count;
	int ret = 0;

	if (new_extents == 0)
		return 0;

	info = btrfs_search_free_space_info(trans, block_group, path, 1);
	if (IS_ERR(info))
		return PTR_ERR(info);

	flags = btrfs_free_space_flags(path->nodes[0], info);
	extent_count = btrfs_free_space_extent_count(path->nodes[0], info);

	extent_count += new_extents;
	btrfs_set_free_space_extent_count(path->nodes[0], info, extent_count);
	btrfs_release_path(path);

	if (!(flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
	    extent_count > block_group->bitmap_high_thresh) {
		ret = btrfs_convert_free_space_to_bitmaps(trans, block_group, path);
	} else if ((flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
		   extent_count < block_group->bitmap_low_thresh) {
		ret = btrfs_convert_free_space_to_extents(trans, block_group, path);
	}

	return ret;
}

EXPORT_FOR_TESTS
bool btrfs_free_space_test_bit(struct btrfs_block_group *block_group,
			       struct btrfs_path *path, u64 offset)
{
	struct extent_buffer *leaf;
	struct btrfs_key key;
	u64 found_start, found_end;
	unsigned long ptr, i;

	leaf = path->nodes[0];
	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
	ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);

	found_start = key.objectid;
	found_end = key.objectid + key.offset;
	ASSERT(offset >= found_start && offset < found_end);

	ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
	i = div_u64(offset - found_start,
		    block_group->fs_info->sectorsize);
	return extent_buffer_test_bit(leaf, ptr, i);
}

static void free_space_modify_bits(struct btrfs_trans_handle *trans,
				   struct btrfs_block_group *block_group,
				   struct btrfs_path *path, u64 *start, u64 *size,
				   bool set_bits)
{
	struct btrfs_fs_info *fs_info = block_group->fs_info;
	struct extent_buffer *leaf;
	struct btrfs_key key;
	u64 end = *start + *size;
	u64 found_start, found_end;
	unsigned long ptr, first, last;

	leaf = path->nodes[0];
	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
	ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);

	found_start = key.objectid;
	found_end = key.objectid + key.offset;
	ASSERT(*start >= found_start && *start < found_end);
	ASSERT(end > found_start);

	if (end > found_end)
		end = found_end;

	ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
	first = (*start - found_start) >> fs_info->sectorsize_bits;
	last = (end - found_start) >> fs_info->sectorsize_bits;
	if (set_bits)
		extent_buffer_bitmap_set(leaf, ptr, first, last - first);
	else
		extent_buffer_bitmap_clear(leaf, ptr, first, last - first);
	btrfs_mark_buffer_dirty(trans, leaf);

	*size -= end - *start;
	*start = end;
}

/*
 * We can't use btrfs_next_item() in modify_free_space_bitmap() because
 * btrfs_next_leaf() doesn't get the path for writing. We can forgo the fancy
 * tree walking in btrfs_next_leaf() anyways because we know exactly what we're
 * looking for.
 */
static int free_space_next_bitmap(struct btrfs_trans_handle *trans,
				  struct btrfs_root *root, struct btrfs_path *p)
{
	struct btrfs_key key;

	if (p->slots[0] + 1 < btrfs_header_nritems(p->nodes[0])) {
		p->slots[0]++;
		return 0;
	}

	btrfs_item_key_to_cpu(p->nodes[0], &key, p->slots[0]);
	btrfs_release_path(p);

	key.objectid += key.offset;
	key.type = (u8)-1;
	key.offset = (u64)-1;

	return btrfs_search_prev_slot(trans, root, &key, p, 0, 1);
}

/*
 * If remove is 1, then we are removing free space, thus clearing bits in the
 * bitmap. If remove is 0, then we are adding free space, thus setting bits in
 * the bitmap.
 */
static int modify_free_space_bitmap(struct btrfs_trans_handle *trans,
				    struct btrfs_block_group *block_group,
				    struct btrfs_path *path,
				    u64 start, u64 size, bool remove)
{
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	struct btrfs_key key;
	u64 end = start + size;
	u64 cur_start, cur_size;
	bool prev_bit_set = false;
	bool next_bit_set = false;
	int new_extents;
	int ret;

	/*
	 * Read the bit for the block immediately before the extent of space if
	 * that block is within the block group.
	 */
	if (start > block_group->start) {
		u64 prev_block = start - block_group->fs_info->sectorsize;

		key.objectid = prev_block;
		key.type = (u8)-1;
		key.offset = (u64)-1;

		ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
		if (ret)
			return ret;

		prev_bit_set = btrfs_free_space_test_bit(block_group, path, prev_block);

		/* The previous block may have been in the previous bitmap. */
		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
		if (start >= key.objectid + key.offset) {
			ret = free_space_next_bitmap(trans, root, path);
			if (ret)
				return ret;
		}
	} else {
		key.objectid = start;
		key.type = (u8)-1;
		key.offset = (u64)-1;

		ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
		if (ret)
			return ret;
	}

	/*
	 * Iterate over all of the bitmaps overlapped by the extent of space,
	 * clearing/setting bits as required.
	 */
	cur_start = start;
	cur_size = size;
	while (1) {
		free_space_modify_bits(trans, block_group, path, &cur_start,
				       &cur_size, !remove);
		if (cur_size == 0)
			break;
		ret = free_space_next_bitmap(trans, root, path);
		if (ret)
			return ret;
	}

	/*
	 * Read the bit for the block immediately after the extent of space if
	 * that block is within the block group.
	 */
	if (end < btrfs_block_group_end(block_group)) {
		/* The next block may be in the next bitmap. */
		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
		if (end >= key.objectid + key.offset) {
			ret = free_space_next_bitmap(trans, root, path);
			if (ret)
				return ret;
		}

		next_bit_set = btrfs_free_space_test_bit(block_group, path, end);
	}

	if (remove) {
		new_extents = -1;
		if (prev_bit_set) {
			/* Leftover on the left. */
			new_extents++;
		}
		if (next_bit_set) {
			/* Leftover on the right. */
			new_extents++;
		}
	} else {
		new_extents = 1;
		if (prev_bit_set) {
			/* Merging with neighbor on the left. */
			new_extents--;
		}
		if (next_bit_set) {
			/* Merging with neighbor on the right. */
			new_extents--;
		}
	}

	btrfs_release_path(path);
	return update_free_space_extent_count(trans, block_group, path, new_extents);
}

static int remove_free_space_extent(struct btrfs_trans_handle *trans,
				    struct btrfs_block_group *block_group,
				    struct btrfs_path *path,
				    u64 start, u64 size)
{
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	struct btrfs_key key;
	u64 found_start, found_end;
	u64 end = start + size;
	int new_extents = -1;
	int ret;

	key.objectid = start;
	key.type = (u8)-1;
	key.offset = (u64)-1;

	ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
	if (ret)
		return ret;

	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);

	ASSERT(key.type == BTRFS_FREE_SPACE_EXTENT_KEY);

	found_start = key.objectid;
	found_end = key.objectid + key.offset;
	ASSERT(start >= found_start && end <= found_end);

	/*
	 * Okay, now that we've found the free space extent which contains the
	 * free space that we are removing, there are four cases:
	 *
	 * 1. We're using the whole extent: delete the key we found and
	 * decrement the free space extent count.
	 * 2. We are using part of the extent starting at the beginning: delete
	 * the key we found and insert a new key representing the leftover at
	 * the end. There is no net change in the number of extents.
	 * 3. We are using part of the extent ending at the end: delete the key
	 * we found and insert a new key representing the leftover at the
	 * beginning. There is no net change in the number of extents.
	 * 4. We are using part of the extent in the middle: delete the key we
	 * found and insert two new keys representing the leftovers on each
	 * side. Where we used to have one extent, we now have two, so increment
	 * the extent count. We may need to convert the block group to bitmaps
	 * as a result.
	 */

	/* Delete the existing key (cases 1-4). */
	ret = btrfs_del_item(trans, root, path);
	if (ret)
		return ret;

	/* Add a key for leftovers at the beginning (cases 3 and 4). */
	if (start > found_start) {
		key.objectid = found_start;
		key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
		key.offset = start - found_start;

		btrfs_release_path(path);
		ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
		if (ret)
			return ret;
		new_extents++;
	}

	/* Add a key for leftovers at the end (cases 2 and 4). */
	if (end < found_end) {
		key.objectid = end;
		key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
		key.offset = found_end - end;

		btrfs_release_path(path);
		ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
		if (ret)
			return ret;
		new_extents++;
	}

	btrfs_release_path(path);
	return update_free_space_extent_count(trans, block_group, path, new_extents);
}

static int using_bitmaps(struct btrfs_block_group *bg, struct btrfs_path *path)
{
	struct btrfs_free_space_info *info;
	u32 flags;

	if (bg->using_free_space_bitmaps_cached)
		return bg->using_free_space_bitmaps;

	info = btrfs_search_free_space_info(NULL, bg, path, 0);
	if (IS_ERR(info))
		return PTR_ERR(info);
	flags = btrfs_free_space_flags(path->nodes[0], info);
	btrfs_release_path(path);

	bg->using_free_space_bitmaps = (flags & BTRFS_FREE_SPACE_USING_BITMAPS);
	bg->using_free_space_bitmaps_cached = true;

	return bg->using_free_space_bitmaps;
}

EXPORT_FOR_TESTS
int __btrfs_remove_from_free_space_tree(struct btrfs_trans_handle *trans,
					struct btrfs_block_group *block_group,
					struct btrfs_path *path, u64 start, u64 size)
{
	int ret;

	ret = __add_block_group_free_space(trans, block_group, path);
	if (ret)
		return ret;

	ret = using_bitmaps(block_group, path);
	if (ret < 0)
		return ret;

	if (ret)
		return modify_free_space_bitmap(trans, block_group, path,
						start, size, true);

	return remove_free_space_extent(trans, block_group, path, start, size);
}

int btrfs_remove_from_free_space_tree(struct btrfs_trans_handle *trans,
				      u64 start, u64 size)
{
	struct btrfs_block_group *block_group;
	BTRFS_PATH_AUTO_FREE(path);
	int ret;

	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
		return 0;

	path = btrfs_alloc_path();
	if (unlikely(!path)) {
		ret = -ENOMEM;
		btrfs_abort_transaction(trans, ret);
		return ret;
	}

	block_group = btrfs_lookup_block_group(trans->fs_info, start);
	if (unlikely(!block_group)) {
		DEBUG_WARN("no block group found for start=%llu", start);
		ret = -ENOENT;
		btrfs_abort_transaction(trans, ret);
		return ret;
	}

	mutex_lock(&block_group->free_space_lock);
	ret = __btrfs_remove_from_free_space_tree(trans, block_group, path, start, size);
	mutex_unlock(&block_group->free_space_lock);
	if (ret)
		btrfs_abort_transaction(trans, ret);

	btrfs_put_block_group(block_group);

	return ret;
}

static int add_free_space_extent(struct btrfs_trans_handle *trans,
				 struct btrfs_block_group *block_group,
				 struct btrfs_path *path,
				 u64 start, u64 size)
{
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	struct btrfs_key key, new_key;
	u64 found_start, found_end;
	u64 end = start + size;
	int new_extents = 1;
	int ret;

	/*
	 * We are adding a new extent of free space, but we need to merge
	 * extents. There are four cases here:
	 *
	 * 1. The new extent does not have any immediate neighbors to merge
	 * with: add the new key and increment the free space extent count. We
	 * may need to convert the block group to bitmaps as a result.
	 * 2. The new extent has an immediate neighbor before it: remove the
	 * previous key and insert a new key combining both of them. There is no
	 * net change in the number of extents.
	 * 3. The new extent has an immediate neighbor after it: remove the next
	 * key and insert a new key combining both of them. There is no net
	 * change in the number of extents.
	 * 4. The new extent has immediate neighbors on both sides: remove both
	 * of the keys and insert a new key combining all of them. Where we used
	 * to have two extents, we now have one, so decrement the extent count.
	 */

	new_key.objectid = start;
	new_key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
	new_key.offset = size;

	/* Search for a neighbor on the left. */
	if (start == block_group->start)
		goto right;
	key.objectid = start - 1;
	key.type = (u8)-1;
	key.offset = (u64)-1;

	ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
	if (ret)
		return ret;

	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);

	if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
		ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
		btrfs_release_path(path);
		goto right;
	}

	found_start = key.objectid;
	found_end = key.objectid + key.offset;
	ASSERT(found_start >= block_group->start &&
	       found_end > block_group->start);
	ASSERT(found_start < start && found_end <= start);

	/*
	 * Delete the neighbor on the left and absorb it into the new key (cases
	 * 2 and 4).
	 */
	if (found_end == start) {
		ret = btrfs_del_item(trans, root, path);
		if (ret)
			return ret;
		new_key.objectid = found_start;
		new_key.offset += key.offset;
		new_extents--;
	}
	btrfs_release_path(path);

right:
	/* Search for a neighbor on the right. */
	if (end == btrfs_block_group_end(block_group))
		goto insert;
	key.objectid = end;
	key.type = (u8)-1;
	key.offset = (u64)-1;

	ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
	if (ret)
		return ret;

	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);

	if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
		ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
		btrfs_release_path(path);
		goto insert;
	}

	found_start = key.objectid;
	found_end = key.objectid + key.offset;
	ASSERT(found_start >= block_group->start &&
	       found_end > block_group->start);
	ASSERT((found_start < start && found_end <= start) ||
	       (found_start >= end && found_end > end));

	/*
	 * Delete the neighbor on the right and absorb it into the new key
	 * (cases 3 and 4).
	 */
	if (found_start == end) {
		ret = btrfs_del_item(trans, root, path);
		if (ret)
			return ret;
		new_key.offset += key.offset;
		new_extents--;
	}
	btrfs_release_path(path);

insert:
	/* Insert the new key (cases 1-4). */
	ret = btrfs_insert_empty_item(trans, root, path, &new_key, 0);
	if (ret)
		return ret;

	btrfs_release_path(path);
	return update_free_space_extent_count(trans, block_group, path, new_extents);
}

EXPORT_FOR_TESTS
int __btrfs_add_to_free_space_tree(struct btrfs_trans_handle *trans,
				   struct btrfs_block_group *block_group,
				   struct btrfs_path *path, u64 start, u64 size)
{
	int ret;

	ret = __add_block_group_free_space(trans, block_group, path);
	if (ret)
		return ret;

	ret = using_bitmaps(block_group, path);
	if (ret < 0)
		return ret;

	if (ret)
		return modify_free_space_bitmap(trans, block_group, path,
						start, size, false);

	return add_free_space_extent(trans, block_group, path, start, size);
}

int btrfs_add_to_free_space_tree(struct btrfs_trans_handle *trans,
				 u64 start, u64 size)
{
	struct btrfs_block_group *block_group;
	BTRFS_PATH_AUTO_FREE(path);
	int ret;

	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
		return 0;

	path = btrfs_alloc_path();
	if (unlikely(!path)) {
		ret = -ENOMEM;
		btrfs_abort_transaction(trans, ret);
		return ret;
	}

	block_group = btrfs_lookup_block_group(trans->fs_info, start);
	if (unlikely(!block_group)) {
		DEBUG_WARN("no block group found for start=%llu", start);
		ret = -ENOENT;
		btrfs_abort_transaction(trans, ret);
		return ret;
	}

	mutex_lock(&block_group->free_space_lock);
	ret = __btrfs_add_to_free_space_tree(trans, block_group, path, start, size);
	mutex_unlock(&block_group->free_space_lock);
	if (ret)
		btrfs_abort_transaction(trans, ret);

	btrfs_put_block_group(block_group);

	return ret;
}

/*
 * Populate the free space tree by walking the extent tree. Operations on the
 * extent tree that happen as a result of writes to the free space tree will go
 * through the normal add/remove hooks.
 */
static int populate_free_space_tree(struct btrfs_trans_handle *trans,
				    struct btrfs_block_group *block_group)
{
	struct btrfs_root *extent_root;
	BTRFS_PATH_AUTO_FREE(path);
	BTRFS_PATH_AUTO_FREE(path2);
	struct btrfs_key key;
	u64 start, end;
	int ret;

	path = btrfs_alloc_path();
	if (!path)
		return -ENOMEM;

	path2 = btrfs_alloc_path();
	if (!path2)
		return -ENOMEM;

	path->reada = READA_FORWARD;

	ret = add_new_free_space_info(trans, block_group, path2);
	if (ret)
		return ret;

	extent_root = btrfs_extent_root(trans->fs_info, block_group->start);
	if (unlikely(!extent_root)) {
		btrfs_err(trans->fs_info,
			  "missing extent root for block group at offset %llu",
			  block_group->start);
		return -EUCLEAN;
	}

	mutex_lock(&block_group->free_space_lock);

	/*
	 * Iterate through all of the extent and metadata items in this block
	 * group, adding the free space between them and the free space at the
	 * end. Note that EXTENT_ITEM and METADATA_ITEM are less than
	 * BLOCK_GROUP_ITEM, so an extent may precede the block group that it's
	 * contained in.
	 */
	key.objectid = block_group->start;
	key.type = BTRFS_EXTENT_ITEM_KEY;
	key.offset = 0;

	ret = btrfs_search_slot_for_read(extent_root, &key, path, 1, 0);
	if (ret < 0)
		goto out_locked;
	/*
	 * If ret is 1 (no key found), it means this is an empty block group,
	 * without any extents allocated from it and there's no block group
	 * item (key BTRFS_BLOCK_GROUP_ITEM_KEY) located in the extent tree
	 * because we are using the block group tree feature (so block group
	 * items are stored in the block group tree) or this is a new block
	 * group created in the current transaction and its block group item
	 * was not yet inserted in the extent tree (that happens in
	 * btrfs_create_pending_block_groups() -> insert_block_group_item()).
	 * It also means there are no extents allocated for block groups with a
	 * start offset beyond this block group's end offset (this is the last,
	 * highest, block group).
	 */
	start = block_group->start;
	end = btrfs_block_group_end(block_group);
	while (ret == 0) {
		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);

		if (key.type == BTRFS_EXTENT_ITEM_KEY ||
		    key.type == BTRFS_METADATA_ITEM_KEY) {
			if (key.objectid >= end)
				break;

			if (start < key.objectid) {
				ret = __btrfs_add_to_free_space_tree(trans,
								     block_group,
								     path2, start,
								     key.objectid -
								     start);
				if (ret)
					goto out_locked;
			}
			start = key.objectid;
			if (key.type == BTRFS_METADATA_ITEM_KEY)
				start += trans->fs_info->nodesize;
			else
				start += key.offset;
		} else if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
			if (key.objectid != block_group->start)
				break;
		}

		ret = btrfs_next_item(extent_root, path);
		if (ret < 0)
			goto out_locked;
	}
	if (start < end) {
		ret = __btrfs_add_to_free_space_tree(trans, block_group, path2,
						     start, end - start);
		if (ret)
			goto out_locked;
	}

	ret = 0;
out_locked:
	mutex_unlock(&block_group->free_space_lock);

	return ret;
}

int btrfs_create_free_space_tree(struct btrfs_fs_info *fs_info)
{
	struct btrfs_trans_handle *trans;
	struct btrfs_root *tree_root = fs_info->tree_root;
	struct btrfs_root *free_space_root;
	struct btrfs_block_group *block_group;
	struct rb_node *node;
	int ret;

	trans = btrfs_start_transaction(tree_root, 0);
	if (IS_ERR(trans))
		return PTR_ERR(trans);

	set_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
	set_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
	free_space_root = btrfs_create_tree(trans,
					    BTRFS_FREE_SPACE_TREE_OBJECTID);
	if (IS_ERR(free_space_root)) {
		ret = PTR_ERR(free_space_root);
		btrfs_abort_transaction(trans, ret);
		btrfs_end_transaction(trans);
		goto out_clear;
	}
	ret = btrfs_global_root_insert(free_space_root);
	if (unlikely(ret)) {
		btrfs_put_root(free_space_root);
		btrfs_abort_transaction(trans, ret);
		btrfs_end_transaction(trans);
		goto out_clear;
	}

	node = rb_first_cached(&fs_info->block_group_cache_tree);
	while (node) {
		block_group = rb_entry(node, struct btrfs_block_group,
				       cache_node);
		ret = populate_free_space_tree(trans, block_group);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			btrfs_end_transaction(trans);
			goto out_clear;
		}
		node = rb_next(node);
	}

	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE);
	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);
	clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
	ret = btrfs_commit_transaction(trans);

	/*
	 * Now that we've committed the transaction any reading of our commit
	 * root will be safe, so we can cache from the free space tree now.
	 */
	clear_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
	return ret;

out_clear:
	clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
	clear_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
	return ret;
}

static int clear_free_space_tree(struct btrfs_trans_handle *trans,
				 struct btrfs_root *root)
{
	BTRFS_PATH_AUTO_FREE(path);
	struct btrfs_key key;
	struct rb_node *node;
	int nr;
	int ret;

	path = btrfs_alloc_path();
	if (!path)
		return -ENOMEM;

	key.objectid = 0;
	key.type = 0;
	key.offset = 0;

	while (1) {
		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
		if (ret < 0)
			return ret;

		nr = btrfs_header_nritems(path->nodes[0]);
		if (!nr)
			break;

		path->slots[0] = 0;
		ret = btrfs_del_items(trans, root, path, 0, nr);
		if (ret)
			return ret;

		btrfs_release_path(path);
	}

	node = rb_first_cached(&trans->fs_info->block_group_cache_tree);
	while (node) {
		struct btrfs_block_group *bg;

		bg = rb_entry(node, struct btrfs_block_group, cache_node);
		clear_bit(BLOCK_GROUP_FLAG_FREE_SPACE_ADDED, &bg->runtime_flags);
		node = rb_next(node);
		cond_resched();
	}

	return 0;
}

int btrfs_delete_free_space_tree(struct btrfs_fs_info *fs_info)
{
	struct btrfs_trans_handle *trans;
	struct btrfs_root *tree_root = fs_info->tree_root;
	struct btrfs_key key = {
		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
		.type = BTRFS_ROOT_ITEM_KEY,
		.offset = 0,
	};
	struct btrfs_root *free_space_root = btrfs_global_root(fs_info, &key);
	int ret;

	trans = btrfs_start_transaction(tree_root, 0);
	if (IS_ERR(trans))
		return PTR_ERR(trans);

	btrfs_clear_fs_compat_ro(fs_info, FREE_SPACE_TREE);
	btrfs_clear_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);

	ret = clear_free_space_tree(trans, free_space_root);
	if (unlikely(ret)) {
		btrfs_abort_transaction(trans, ret);
		btrfs_end_transaction(trans);
		return ret;
	}

	ret = btrfs_del_root(trans, &free_space_root->root_key);
	if (unlikely(ret)) {
		btrfs_abort_transaction(trans, ret);
		btrfs_end_transaction(trans);
		return ret;
	}

	btrfs_global_root_delete(free_space_root);

	spin_lock(&fs_info->trans_lock);
	list_del(&free_space_root->dirty_list);
	spin_unlock(&fs_info->trans_lock);

	btrfs_tree_lock(free_space_root->node);
	btrfs_clear_buffer_dirty(trans, free_space_root->node);
	btrfs_tree_unlock(free_space_root->node);
	ret = btrfs_free_tree_block(trans, btrfs_root_id(free_space_root),
				    free_space_root->node, 0, 1);
	btrfs_put_root(free_space_root);
	if (unlikely(ret < 0)) {
		btrfs_abort_transaction(trans, ret);
		btrfs_end_transaction(trans);
		return ret;
	}

	return btrfs_commit_transaction(trans);
}

int btrfs_rebuild_free_space_tree(struct btrfs_fs_info *fs_info)
{
	struct btrfs_trans_handle *trans;
	struct btrfs_key key = {
		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
		.type = BTRFS_ROOT_ITEM_KEY,
		.offset = 0,
	};
	struct btrfs_root *free_space_root = btrfs_global_root(fs_info, &key);
	struct rb_node *node;
	int ret;

	trans = btrfs_start_transaction(free_space_root, 1);
	if (IS_ERR(trans))
		return PTR_ERR(trans);

	set_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
	set_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);

	ret = clear_free_space_tree(trans, free_space_root);
	if (unlikely(ret)) {
		btrfs_abort_transaction(trans, ret);
		btrfs_end_transaction(trans);
		return ret;
	}

	node = rb_first_cached(&fs_info->block_group_cache_tree);
	while (node) {
		struct btrfs_block_group *block_group;

		block_group = rb_entry(node, struct btrfs_block_group,
				       cache_node);

		if (test_bit(BLOCK_GROUP_FLAG_FREE_SPACE_ADDED,
			     &block_group->runtime_flags))
			goto next;

		ret = populate_free_space_tree(trans, block_group);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			btrfs_end_transaction(trans);
			return ret;
		}
next:
		if (btrfs_should_end_transaction(trans)) {
			btrfs_end_transaction(trans);
			trans = btrfs_start_transaction(free_space_root, 1);
			if (IS_ERR(trans))
				return PTR_ERR(trans);
		}
		node = rb_next(node);
	}

	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE);
	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);
	clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);

	ret = btrfs_commit_transaction(trans);
	clear_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
	return ret;
}

static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
					struct btrfs_block_group *block_group,
					struct btrfs_path *path)
{
	bool own_path = false;
	int ret;

	if (!test_and_clear_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE,
				&block_group->runtime_flags))
		return 0;

	/*
	 * While rebuilding the free space tree we may allocate new metadata
	 * block groups while modifying the free space tree.
	 *
	 * Because during the rebuild (at btrfs_rebuild_free_space_tree()) we
	 * can use multiple transactions, every time btrfs_end_transaction() is
	 * called at btrfs_rebuild_free_space_tree() we finish the creation of
	 * new block groups by calling btrfs_create_pending_block_groups(), and
	 * that in turn calls us, through btrfs_add_block_group_free_space(),
	 * to add a free space info item and a free space extent item for the
	 * block group.
	 *
	 * Then later btrfs_rebuild_free_space_tree() may find such new block
	 * groups and processes them with populate_free_space_tree(), which can
	 * fail with EEXIST since there are already items for the block group in
	 * the free space tree. Notice that we say "may find" because a new
	 * block group may be added to the block groups rbtree in a node before
	 * or after the block group currently being processed by the rebuild
	 * process. So signal the rebuild process to skip such new block groups
	 * if it finds them.
	 */
	set_bit(BLOCK_GROUP_FLAG_FREE_SPACE_ADDED, &block_group->runtime_flags);

	if (!path) {
		path = btrfs_alloc_path();
		if (unlikely(!path)) {
			btrfs_abort_transaction(trans, -ENOMEM);
			return -ENOMEM;
		}
		own_path = true;
	}

	ret = add_new_free_space_info(trans, block_group, path);
	if (unlikely(ret)) {
		btrfs_abort_transaction(trans, ret);
		goto out;
	}

	ret = __btrfs_add_to_free_space_tree(trans, block_group, path,
					     block_group->start, block_group->length);
	if (ret)
		btrfs_abort_transaction(trans, ret);

out:
	if (own_path)
		btrfs_free_path(path);

	return ret;
}

int btrfs_add_block_group_free_space(struct btrfs_trans_handle *trans,
				     struct btrfs_block_group *block_group)
{
	int ret;

	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
		return 0;

	mutex_lock(&block_group->free_space_lock);
	ret = __add_block_group_free_space(trans, block_group, NULL);
	mutex_unlock(&block_group->free_space_lock);
	return ret;
}

int btrfs_remove_block_group_free_space(struct btrfs_trans_handle *trans,
					struct btrfs_block_group *block_group)
{
	struct btrfs_root *root = btrfs_free_space_root(block_group);
	BTRFS_PATH_AUTO_FREE(path);
	struct btrfs_key key, found_key;
	struct extent_buffer *leaf;
	u64 start, end;
	bool done = false;
	int nr;
	int ret;

	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
		return 0;

	if (test_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &block_group->runtime_flags)) {
		/* We never added this block group to the free space tree. */
		return 0;
	}

	path = btrfs_alloc_path();
	if (unlikely(!path)) {
		ret = -ENOMEM;
		btrfs_abort_transaction(trans, ret);
		return ret;
	}

	start = block_group->start;
	end = btrfs_block_group_end(block_group);

	key.objectid = end - 1;
	key.type = (u8)-1;
	key.offset = (u64)-1;

	while (!done) {
		ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			return ret;
		}

		leaf = path->nodes[0];
		nr = 0;
		path->slots[0]++;
		while (path->slots[0] > 0) {
			btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);

			if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
				ASSERT(found_key.objectid == block_group->start);
				ASSERT(found_key.offset == block_group->length);
				done = true;
				nr++;
				path->slots[0]--;
				break;
			} else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY ||
				   found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
				ASSERT(found_key.objectid >= start);
				ASSERT(found_key.objectid < end);
				ASSERT(found_key.objectid + found_key.offset <= end);
				nr++;
				path->slots[0]--;
			} else {
				btrfs_err(trans->fs_info, "unexpected free space tree key type %u",
					  found_key.type);
				ret = -EUCLEAN;
				btrfs_abort_transaction(trans, ret);
				return ret;
			}
		}

		ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
		if (unlikely(ret)) {
			btrfs_abort_transaction(trans, ret);
			return ret;
		}
		btrfs_release_path(path);
	}

	return 0;
}

static int validate_free_space_key(struct btrfs_block_group *block_group,
				   const struct btrfs_key *key, u8 expected_type)
{
	const u64 end = btrfs_block_group_end(block_group);

	if (unlikely(key->type != expected_type)) {
		btrfs_err(block_group->fs_info,
			  "block group %llu has unexpected free space key type %u, expected %u",
			  block_group->start, key->type, expected_type);
		return -EUCLEAN;
	}

	if (unlikely(key->objectid + key->offset > end)) {
		btrfs_err(block_group->fs_info,
			  "block group %llu has invalid free space key (%llu %u %llu)",
			  block_group->start, key->objectid, key->type,
			  key->offset);
		return -EUCLEAN;
	}

	return 0;
}

static int load_free_space_bitmaps(struct btrfs_caching_control *caching_ctl,
				   struct btrfs_path *path,
				   u32 expected_extent_count)
{
	struct btrfs_block_group *block_group = caching_ctl->block_group;
	struct btrfs_fs_info *fs_info = block_group->fs_info;
	struct btrfs_root *root;
	struct btrfs_key key;
	bool prev_bit_set = false;
	/* Initialize to silence GCC. */
	u64 extent_start = 0;
	const u64 end = btrfs_block_group_end(block_group);
	u64 offset;
	u64 total_found = 0;
	u32 extent_count = 0;
	int ret;

	root = btrfs_free_space_root(block_group);

	while (1) {
		ret = btrfs_next_item(root, path);
		if (ret < 0)
			return ret;
		if (ret)
			break;

		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);

		if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
			break;

		ret = validate_free_space_key(block_group, &key, BTRFS_FREE_SPACE_BITMAP_KEY);
		if (unlikely(ret))
			return ret;

		offset = key.objectid;
		while (offset < key.objectid + key.offset) {
			bool bit_set;

			bit_set = btrfs_free_space_test_bit(block_group, path, offset);
			if (!prev_bit_set && bit_set) {
				extent_start = offset;
			} else if (prev_bit_set && !bit_set) {
				u64 space_added;

				ret = btrfs_add_new_free_space(block_group,
							       extent_start,
							       offset,
							       &space_added);
				if (ret)
					return ret;
				total_found += space_added;
				if (total_found > CACHING_CTL_WAKE_UP) {
					total_found = 0;
					wake_up(&caching_ctl->wait);
				}
				extent_count++;
			}
			prev_bit_set = bit_set;
			offset += fs_info->sectorsize;
		}
	}
	if (prev_bit_set) {
		ret = btrfs_add_new_free_space(block_group, extent_start, end, NULL);
		if (ret)
			return ret;
		extent_count++;
	}

	if (unlikely(extent_count != expected_extent_count)) {
		btrfs_err(fs_info,
			  "incorrect extent count for %llu; counted %u, expected %u",
			  block_group->start, extent_count,
			  expected_extent_count);
		DEBUG_WARN();
		return -EIO;
	}

	return 0;
}

static int load_free_space_extents(struct btrfs_caching_control *caching_ctl,
				   struct btrfs_path *path,
				   u32 expected_extent_count)
{
	struct btrfs_block_group *block_group = caching_ctl->block_group;
	struct btrfs_fs_info *fs_info = block_group->fs_info;
	struct btrfs_root *root;
	struct btrfs_key key;
	u64 total_found = 0;
	u32 extent_count = 0;
	int ret;

	root = btrfs_free_space_root(block_group);

	while (1) {
		u64 space_added;

		ret = btrfs_next_item(root, path);
		if (ret < 0)
			return ret;
		if (ret)
			break;

		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);

		if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
			break;

		ret = validate_free_space_key(block_group, &key, BTRFS_FREE_SPACE_EXTENT_KEY);
		if (unlikely(ret))
			return ret;

		ret = btrfs_add_new_free_space(block_group, key.objectid,
					       key.objectid + key.offset,
					       &space_added);
		if (ret)
			return ret;
		total_found += space_added;
		if (total_found > CACHING_CTL_WAKE_UP) {
			total_found = 0;
			wake_up(&caching_ctl->wait);
		}
		extent_count++;
	}

	if (unlikely(extent_count != expected_extent_count)) {
		btrfs_err(fs_info,
			  "incorrect extent count for %llu; counted %u, expected %u",
			  block_group->start, extent_count,
			  expected_extent_count);
		DEBUG_WARN();
		return -EIO;
	}

	return 0;
}

int btrfs_load_free_space_tree(struct btrfs_caching_control *caching_ctl)
{
	struct btrfs_block_group *block_group;
	struct btrfs_free_space_info *info;
	BTRFS_PATH_AUTO_FREE(path);
	u32 extent_count, flags;

	block_group = caching_ctl->block_group;

	path = btrfs_alloc_path();
	if (!path)
		return -ENOMEM;

	/*
	 * Just like caching_thread() doesn't want to deadlock on the extent
	 * tree, we don't want to deadlock on the free space tree.
	 */
	path->skip_locking = true;
	path->search_commit_root = true;
	path->reada = READA_FORWARD;

	info = btrfs_search_free_space_info(NULL, block_group, path, 0);
	if (IS_ERR(info))
		return PTR_ERR(info);

	extent_count = btrfs_free_space_extent_count(path->nodes[0], info);
	flags = btrfs_free_space_flags(path->nodes[0], info);

	/*
	 * We left path pointing to the free space info item, so now
	 * load_free_space_foo can just iterate through the free space tree from
	 * there.
	 */
	if (flags & BTRFS_FREE_SPACE_USING_BITMAPS)
		return load_free_space_bitmaps(caching_ctl, path, extent_count);
	else
		return load_free_space_extents(caching_ctl, path, extent_count);
}

static int delete_orphan_free_space_entries(struct btrfs_root *fst_root,
					    struct btrfs_path *path,
					    u64 first_bg_bytenr)
{
	struct btrfs_trans_handle *trans;
	int ret;

	trans = btrfs_start_transaction(fst_root, 1);
	if (IS_ERR(trans))
		return PTR_ERR(trans);

	while (true) {
		struct btrfs_key key = { 0 };
		int i;

		ret = btrfs_search_slot(trans, fst_root, &key, path, -1, 1);
		if (ret < 0)
			break;
		ASSERT(ret > 0);
		ret = 0;
		for (i = 0; i < btrfs_header_nritems(path->nodes[0]); i++) {
			btrfs_item_key_to_cpu(path->nodes[0], &key, i);
			if (key.objectid >= first_bg_bytenr) {
				/*
				 * Only break the for() loop and continue to
				 * delete items.
				 */
				break;
			}
		}
		/* No items to delete, finished. */
		if (i == 0)
			break;

		ret = btrfs_del_items(trans, fst_root, path, 0, i);
		if (ret < 0)
			break;
		btrfs_release_path(path);
	}
	btrfs_release_path(path);
	btrfs_end_transaction(trans);
	if (ret == 0)
		btrfs_info(fst_root->fs_info, "deleted orphan free space tree entries");
	return ret;
}

/* Remove any free space entry before the first block group. */
int btrfs_delete_orphan_free_space_entries(struct btrfs_fs_info *fs_info)
{
	BTRFS_PATH_AUTO_RELEASE(path);
	struct btrfs_key key = {
		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
		.type = BTRFS_ROOT_ITEM_KEY,
		.offset = 0,
	};
	struct btrfs_root *root;
	struct btrfs_block_group *bg;
	u64 first_bg_bytenr;
	int ret;

	/*
	 * Extent tree v2 has multiple global roots based on the block group.
	 * This means we cannot easily grab the global free space tree and locate
	 * orphan items.  Furthermore this is still experimental, all users
	 * should use the latest btrfs-progs anyway.
	 */
	if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
		return 0;
	if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
		return 0;
	root = btrfs_global_root(fs_info, &key);
	if (!root)
		return 0;

	key.objectid = 0;
	key.type = 0;
	key.offset = 0;

	bg = btrfs_lookup_first_block_group(fs_info, 0);
	if (unlikely(!bg)) {
		btrfs_err(fs_info, "no block group found");
		return -EUCLEAN;
	}
	first_bg_bytenr = bg->start;
	btrfs_put_block_group(bg);

	ret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
	if (ret < 0)
		return ret;
	/* There should not be an all-zero key in fst. */
	ASSERT(ret > 0);

	/* Empty free space tree. */
	if (path.slots[0] >= btrfs_header_nritems(path.nodes[0]))
		return 0;

	btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
	if (key.objectid >= first_bg_bytenr)
		return 0;
	btrfs_release_path(&path);
	return delete_orphan_free_space_entries(root, &path, first_bg_bytenr);
}