Re: [PATCH RFC] generic/730: ensure EIO after device delete

["Darrick J. Wong" <djwong@kernel.org>]

On Wed, Mar 13, 2024 at 04:47:47PM -0700, Boris Burkov wrote:
> This test removes a SCSI debug device out from under a mounted
> filesystem with a (probably) dirty file. This assumes that page cache
> cannot save us from EIO, for a reason that I can't quite explain. In
> fact, this test fails for exactly that reason, at least on btrfs.
> 
> The original patches:
> 
>       https://lore.kernel.org/fstests/20230807112100.GB15405@lst.de/
> 
> refer to this passing on xfs and not btrfs, so I suspect I am missing
> something. With that said, on my machine this actually fails on xfs with
> and without my patch, so this is clearly not enough.
> 
> High level, I am trying to understand what is really the expected
> behavior from a filesystem under this condition and what this test is
> getting at. Of btrfs, ext4, and xfs, only ext4 passes it, while btrfs
> does pass with this additional syncing/cache dropping to nudge it to an
> error.

Does btrfs prefetch pagecache data as soon as a file opens?  Or I guess
it could be that xfs trips an IO error and shuts down, and
xfs_file_read_iter will return EIO after that happens.

--D

> Signed-off-by: Boris Burkov <boris@bur.io>
> ---
>  tests/generic/730 | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/tests/generic/730 b/tests/generic/730
> index 11308cdaa..ca5037c57 100755
> --- a/tests/generic/730
> +++ b/tests/generic/730
> @@ -47,6 +47,9 @@ exec 3< $SCSI_DEBUG_MNT/testfile
>  # delete the scsi debug device while it still has dirty data
>  echo 1 > /sys/block/$(_short_dev $SCSI_DEBUG_DEV)/device/delete
>  
> +sync
> +echo 3 > /proc/sys/vm/drop_caches
> +
>  # try to read from the file, which should give us -EIO
>  cat <&3 > /dev/null
>  
> -- 
> 2.43.0
> 
>