On Tue, Jun 28, 2022 at 01:22:02PM -0700, Darrick J. Wong wrote:
> From: Darrick J. Wong <djwong@xxxxxxxxxx>
>
> Beginning with 5.18, some filesystems support creating large folios for
> the page cache. A system with 64k pages can create 256k folios, which
> means that with the old file size of 1M, the last half of the file is
> completely converted from unwritten to written by page_mkwrite. The
> test encodes a translated version of the xfs_bmap output in the golden
> output, which means that the test now fails on 64k pages. Fixing the
> 64k page case by increasing the file size to 2MB broke fsdax because
> fsdax uses 2MB PMDs, hence 12MB.
>
> Increase the size to prevent this from happening. This may require
> further revision if folios get larger or fsdax starts supporting PMDs
> that are larger than 2MB.
>
> Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
> ---
> tests/xfs/166 | 19 +++++++++++++++----
> 1 file changed, 15 insertions(+), 4 deletions(-)
>
>
> diff --git a/tests/xfs/166 b/tests/xfs/166
> index 42379961..d45dc5e8 100755
> --- a/tests/xfs/166
> +++ b/tests/xfs/166
> @@ -16,12 +16,12 @@ _begin_fstest rw metadata auto quick
> # the others are unwritten.
> _filter_blocks()
> {
> - $AWK_PROG '
> + $AWK_PROG -v file_size=$FILE_SIZE '
> /^ +[0-9]/ {
> if (!written_size) {
> written_size = $6
> - unwritten1 = ((1048576/512) / 2) - written_size
> - unwritten2 = ((1048576/512) / 2) - 2 * written_size
> + unwritten1 = ((file_size/512) / 2) - written_size
> + unwritten2 = ((file_size/512) / 2) - 2 * written_size
> }
>
> # is the extent unwritten?
> @@ -58,7 +58,18 @@ _scratch_mount
>
> TEST_FILE=$SCRATCH_MNT/test_file
> TEST_PROG=$here/src/unwritten_mmap
> -FILE_SIZE=1048576
> +
> +# Beginning with 5.18, some filesystems support creating large folios for the
> +# page cache. A system with 64k pages can create 256k folios, which means
> +# that with the old file size of 1M, the last half of the file is completely
> +# converted from unwritten to written by page_mkwrite. The test will fail on
> +# the golden output when this happens, so increase the size from the original
> +# 1MB file size to at least (6 * 256k == 1.5MB) prevent this from happening.
> +#
> +# However, increasing the file size to around 2MB causes regressions when fsdax
> +# is enabled because fsdax will try to use PMD entries for the mappings. Hence
> +# we need to set the file size to (6 * 2MB == 12MB) to cover all cases.
> +FILE_SIZE=$((12 * 1048576))
LGTM.
Reviewed-by: Dave Chinner <dchinner@xxxxxxxxxx>
--
Dave Chinner
david@xxxxxxxxxxxxx
