On Fri, Sep 01, 2023 at 09:56:29PM -0700, Darrick J. Wong wrote:
> On Sat, Sep 02, 2023 at 03:08:02AM +0800, Zorro Lang wrote:
> > On Tue, Aug 29, 2023 at 04:03:49PM -0700, Darrick J. Wong wrote:
> > > From: Darrick J. Wong <djwong@xxxxxxxxxx>
> > >
> > > The write invalidation code in iomap can only be triggered for writes
> > > that span multiple folios. If the kernel reports a huge page size,
> > > scale up the write size.
> > >
> > > Signed-off-by: Darrick J. Wong <djwong@xxxxxxxxxx>
> > > ---
> > > tests/xfs/559 | 29 ++++++++++++++++++++++++++++-
> > > 1 file changed, 28 insertions(+), 1 deletion(-)
> > >
> > >
> > > diff --git a/tests/xfs/559 b/tests/xfs/559
> > > index cffe5045a5..64fc16ebfd 100755
> > > --- a/tests/xfs/559
> > > +++ b/tests/xfs/559
> > > @@ -42,11 +42,38 @@ $XFS_IO_PROG -c 'chattr -x' $SCRATCH_MNT &> $seqres.full
> > > _require_pagecache_access $SCRATCH_MNT
> > >
> > > blocks=10
> > > -blksz=$(_get_page_size)
> > > +
> > > +# If this kernel advertises huge page support, it's possible that it could be
> > > +# using large folios for the page cache writes. It is necessary to write
> > > +# multiple folios (large or regular) to triggering the write invalidation,
> > > +# so we'll scale the test write size accordingly.
> > > +blksz=$(_get_hugepagesize)
> >
> > Isn't _require_hugepages needed if _get_hugepagesize is used?
>
> Nope -- if the kernel doesn't support hugepages, then _get_hugepagesize
> returns the empty string...
>
> > Thanks,
> > Zorro
> >
> > > +base_pagesize=$(_get_page_size)
> > > +test -z "$blksz" && blksz=${base_pagesize}
>
> ...and this line will substitute in the base page size as the block size.
Oh yes, you catch it at here :) OK, so the whole patchset is good to me,
Reviewed-by: Zorro Lang <zlang@xxxxxxxxxx>
>
> --D
>
> > > filesz=$((blocks * blksz))
> > > dirty_offset=$(( filesz - 1 ))
> > > write_len=$(( ( (blocks - 1) * blksz) + 1 ))
> > >
> > > +# The write invalidation that we're testing below can only occur as part of
> > > +# a single large write. The kernel limits writes to one base page less than
> > > +# 2GiB to prevent lengthy IOs and integer overflows. If the block size is so
> > > +# huge (e.g. 512M huge pages on arm64) that we'd exceed that, reduce the number
> > > +# of blocks to get us under the limit.
> > > +max_writesize=$((2147483647 - base_pagesize))
> > > +if ((write_len > max_writesize)); then
> > > + blocks=$(( ( (max_writesize - 1) / blksz) + 1))
> > > + # We need at least three blocks in the file to test invalidation
> > > + # between writes to multiple folios. If we drop below that,
> > > + # reconfigure ourselves with base pages and hope for the best.
> > > + if ((blocks < 3)); then
> > > + blksz=$base_pagesize
> > > + blocks=10
> > > + fi
> > > + filesz=$((blocks * blksz))
> > > + dirty_offset=$(( filesz - 1 ))
> > > + write_len=$(( ( (blocks - 1) * blksz) + 1 ))
> > > +fi
> > > +
> > > # Create a large file with a large unwritten range.
> > > $XFS_IO_PROG -f -c "falloc 0 $filesz" $SCRATCH_MNT/file >> $seqres.full
> > >
> > >
> >
>
