On Wed, May 13, 2020 at 12:48:03PM +0300, Dan Carpenter wrote:
> Originally this function used to always return -EFSCORRUPTED on error
> but now we're trying to return more informative error codes.
> Unfortunately, there was one error path missed. If this kmem_alloc()
> allocation fails then we need to return -ENOMEM instead of success.
>
> Fixes: f20192991d79 ("xfs: simplify inode flush error handling")
This logic predates that patch, and I think it may be by design. Inode
cluster flushing is an optimization to flush other dirty inodes in the
same cluster we're about to queue for writeback. If the cluster flush
fails due to an operational error such as memory allocation failure, we
don't want to report an error because that would shutdown the fs when
otherwise the side effect would be that the other inodes in the cluster
would be flushed individually. This is distinct from failing to flush a
particular inode due to corruption, which is a fatal filesystem error.
Brian
> Signed-off-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
> ---
> fs/xfs/xfs_inode.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
> index ab31a5dec7aab..63aeda7cbafb0 100644
> --- a/fs/xfs/xfs_inode.c
> +++ b/fs/xfs/xfs_inode.c
> @@ -3505,8 +3505,10 @@ xfs_iflush_cluster(
>
> cilist_size = igeo->inodes_per_cluster * sizeof(struct xfs_inode *);
> cilist = kmem_alloc(cilist_size, KM_MAYFAIL|KM_NOFS);
> - if (!cilist)
> + if (!cilist) {
> + error = -ENOMEM;
> goto out_put;
> + }
>
> mask = ~(igeo->inodes_per_cluster - 1);
> first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
> --
> 2.26.2
>
