On 2018/9/13 上午4:49, damenly.su@xxxxxxxxx wrote:
> From: Su Yue <suy.fnst@xxxxxxxxxxxxxx>
>
> In check_fs_roots_lowmem(), we do search and follow the resulted path
> to call check_fs_root(), then call btrfs_next_item() to check next
> root.
> However, if repair is enabled, the root tree can be cowed, the
> existed path can cause strange errors.
>
> Solution:
> If repair, save the key before calling check_fs_root,
> search the saved key again before checking next root.
Both reason and solution looks good.
>
> Signed-off-by: Su Yue <suy.fnst@xxxxxxxxxxxxxx>
> ---
> check/mode-lowmem.c | 15 +++++++++++++++
> 1 file changed, 15 insertions(+)
>
> diff --git a/check/mode-lowmem.c b/check/mode-lowmem.c
> index 89a304bbdd69..8fc9edab1d66 100644
> --- a/check/mode-lowmem.c
> +++ b/check/mode-lowmem.c
> @@ -4967,9 +4967,13 @@ int check_fs_roots_lowmem(struct btrfs_fs_info *fs_info)
> }
>
> while (1) {
> + struct btrfs_key saved_key;
> +
> node = path.nodes[0];
> slot = path.slots[0];
> btrfs_item_key_to_cpu(node, &key, slot);
> + if (repair)
> + saved_key = key;
> if (key.objectid > BTRFS_LAST_FREE_OBJECTID)
> goto out;
> if (key.type == BTRFS_ROOT_ITEM_KEY &&
> @@ -5000,6 +5004,17 @@ int check_fs_roots_lowmem(struct btrfs_fs_info *fs_info)
> err |= ret;
> }
> next:
> + /*
> + * Since root tree can be cowed during repair,
> + * here search the saved key again.
> + */
> + if (repair) {
> + btrfs_release_path(&path);
> + ret = btrfs_search_slot(NULL, fs_info->tree_root,
> + &saved_key, &path, 0, 0);
> + /* Repair never deletes trees, search must succeed. */
> + BUG_ON(ret);
But this doesn't look good to me.
Your assumption here is valid (at least for now), but it's possible that
some tree blocks get corrupted in a large root tree, and in that case,
we could still read part of the root tree, but btrfs_search_slot() could
still return -EIO for certain search key.
So I still prefer to do some error handling other than BUG_ON(ret).
Thanks,
Qu
> + }
> ret = btrfs_next_item(tree_root, &path);
> if (ret > 0)
> goto out;
>
