linux/fs/btrfs/block-group.c

// SPDX-License-Identifier: GPL-2.0

#include "ctree.h"
#include "block-group.h"
#include "space-info.h"

void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
{
	atomic_inc(&cache->count);
}

void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
{
	if (atomic_dec_and_test(&cache->count)) {
		WARN_ON(cache->pinned > 0);
		WARN_ON(cache->reserved > 0);

		/*
		 * If not empty, someone is still holding mutex of
		 * full_stripe_lock, which can only be released by caller.
		 * And it will definitely cause use-after-free when caller
		 * tries to release full stripe lock.
		 *
		 * No better way to resolve, but only to warn.
		 */
		WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
		kfree(cache->free_space_ctl);
		kfree(cache);
	}
}

/*
 * This will return the block group at or after bytenr if contains is 0, else
 * it will return the block group that contains the bytenr
 */
static struct btrfs_block_group_cache *block_group_cache_tree_search(
		struct btrfs_fs_info *info, u64 bytenr, int contains)
{
	struct btrfs_block_group_cache *cache, *ret = NULL;
	struct rb_node *n;
	u64 end, start;

	spin_lock(&info->block_group_cache_lock);
	n = info->block_group_cache_tree.rb_node;

	while (n) {
		cache = rb_entry(n, struct btrfs_block_group_cache,
				 cache_node);
		end = cache->key.objectid + cache->key.offset - 1;
		start = cache->key.objectid;

		if (bytenr < start) {
			if (!contains && (!ret || start < ret->key.objectid))
				ret = cache;
			n = n->rb_left;
		} else if (bytenr > start) {
			if (contains && bytenr <= end) {
				ret = cache;
				break;
			}
			n = n->rb_right;
		} else {
			ret = cache;
			break;
		}
	}
	if (ret) {
		btrfs_get_block_group(ret);
		if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
			info->first_logical_byte = ret->key.objectid;
	}
	spin_unlock(&info->block_group_cache_lock);

	return ret;
}

/*
 * Return the block group that starts at or after bytenr
 */
struct btrfs_block_group_cache *btrfs_lookup_first_block_group(
		struct btrfs_fs_info *info, u64 bytenr)
{
	return block_group_cache_tree_search(info, bytenr, 0);
}

/*
 * Return the block group that contains the given bytenr
 */
struct btrfs_block_group_cache *btrfs_lookup_block_group(
		struct btrfs_fs_info *info, u64 bytenr)
{
	return block_group_cache_tree_search(info, bytenr, 1);
}

struct btrfs_block_group_cache *btrfs_next_block_group(
		struct btrfs_block_group_cache *cache)
{
	struct btrfs_fs_info *fs_info = cache->fs_info;
	struct rb_node *node;

	spin_lock(&fs_info->block_group_cache_lock);

	/* If our block group was removed, we need a full search. */
	if (RB_EMPTY_NODE(&cache->cache_node)) {
		const u64 next_bytenr = cache->key.objectid + cache->key.offset;

		spin_unlock(&fs_info->block_group_cache_lock);
		btrfs_put_block_group(cache);
		cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
	}
	node = rb_next(&cache->cache_node);
	btrfs_put_block_group(cache);
	if (node) {
		cache = rb_entry(node, struct btrfs_block_group_cache,
				 cache_node);
		btrfs_get_block_group(cache);
	} else
		cache = NULL;
	spin_unlock(&fs_info->block_group_cache_lock);
	return cache;
}

bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
{
	struct btrfs_block_group_cache *bg;
	bool ret = true;

	bg = btrfs_lookup_block_group(fs_info, bytenr);
	if (!bg)
		return false;

	spin_lock(&bg->lock);
	if (bg->ro)
		ret = false;
	else
		atomic_inc(&bg->nocow_writers);
	spin_unlock(&bg->lock);

	/* No put on block group, done by btrfs_dec_nocow_writers */
	if (!ret)
		btrfs_put_block_group(bg);

	return ret;
}

void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
{
	struct btrfs_block_group_cache *bg;

	bg = btrfs_lookup_block_group(fs_info, bytenr);
	ASSERT(bg);
	if (atomic_dec_and_test(&bg->nocow_writers))
		wake_up_var(&bg->nocow_writers);
	/*
	 * Once for our lookup and once for the lookup done by a previous call
	 * to btrfs_inc_nocow_writers()
	 */
	btrfs_put_block_group(bg);
	btrfs_put_block_group(bg);
}

void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
{
	wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
}

void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
					const u64 start)
{
	struct btrfs_block_group_cache *bg;

	bg = btrfs_lookup_block_group(fs_info, start);
	ASSERT(bg);
	if (atomic_dec_and_test(&bg->reservations))
		wake_up_var(&bg->reservations);
	btrfs_put_block_group(bg);
}

void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
{
	struct btrfs_space_info *space_info = bg->space_info;

	ASSERT(bg->ro);

	if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
		return;

	/*
	 * Our block group is read only but before we set it to read only,
	 * some task might have had allocated an extent from it already, but it
	 * has not yet created a respective ordered extent (and added it to a
	 * root's list of ordered extents).
	 * Therefore wait for any task currently allocating extents, since the
	 * block group's reservations counter is incremented while a read lock
	 * on the groups' semaphore is held and decremented after releasing
	 * the read access on that semaphore and creating the ordered extent.
	 */
	down_write(&space_info->groups_sem);
	up_write(&space_info->groups_sem);

	wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
}