Re: [PATCH 7/8] generic/371: disable speculative preallocation regressions on XFS

[Allison Henderson <allison.henderson@oracle.com>]

On 7/6/21 5:21 PM, Darrick J. Wong wrote:
> From: Darrick J. Wong <djwong@kernel.org>
> 
> Once in a very long while, the fallocate calls in this test will fail
> due to ENOSPC conditions.  While in theory this test is careful only to
> allocate at most 160M of space from a 256M filesystem, there's a twist
> on XFS: speculative preallocation.
> 
> The first loop in this test is an 80M appending write done in units of
> 4k.  Once the file size hits 64k, XFS will begin speculatively
> preallocating blocks past the end of the file; as the file grows larger,
> so will the speculative preallocation.
> 
> Since the pwrite/rm loop races with the fallocate/rm loop, it's possible
> that the fallocate loop will free that file just before the buffered
> write extends the speculative preallocation out to 160MB.  With fs and
> log overhead, that doesn't leave enough free space to start the 80MB
> fallocate request, which tries to avoid disappointing the caller by
> freeing all speculative preallocations.  That fails if the pwriter
> thread owns the IOLOCK on $testfile1, so fallocate returns ENOSPC and
> the test fails.
> 
> The simple solution here is to disable speculative preallocation by
> setting an extent size hint if the fs is XFS.
> 
Ok, makes sense
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
> ---
>   tests/generic/371 |    8 ++++++++
>   1 file changed, 8 insertions(+)
> 
> 
> diff --git a/tests/generic/371 b/tests/generic/371
> index c94fa85e..a2fdaf7b 100755
> --- a/tests/generic/371
> +++ b/tests/generic/371
> @@ -18,10 +18,18 @@ _begin_fstest auto quick enospc prealloc
>   _supported_fs generic
>   _require_scratch
>   _require_xfs_io_command "falloc"
> +test "$FSTYP" = "xfs" && _require_xfs_io_command "extsize"
>   
>   _scratch_mkfs_sized $((256 * 1024 * 1024)) >> $seqres.full 2>&1
>   _scratch_mount
>   
> +# Disable speculative post-EOF preallocation on XFS, which can grow fast enough
> +# that a racing fallocate then fails.
> +if [ "$FSTYP" = "xfs" ]; then
> +	alloc_sz="$(_get_file_block_size $SCRATCH_MNT)"
> +	$XFS_IO_PROG -c "extsize $alloc_sz" $SCRATCH_MNT >> $seqres.full
> +fi
> +
>   testfile1=$SCRATCH_MNT/testfile1
>   testfile2=$SCRATCH_MNT/testfile2
>   
>