On Wed, Mar 01, 2023 at 12:04:17PM +1100, Dave Chinner wrote:
> On Tue, Feb 28, 2023 at 04:47:01PM -0800, Darrick J. Wong wrote:
> > On Wed, Mar 01, 2023 at 11:17:06AM +1100, Dave Chinner wrote:
> > > From: Dave Chinner <dchinner@xxxxxxxxxx>
> > >
> > > The recent writeback corruption fixes changed the code in
> > > xfs_discard_folio() to calculate a byte range to for punching
> > > delalloc extents. A mistake was made in using round_up(pos) for the
> > > end offset, because when pos points at the first byte of a block, it
> > > does not get rounded up to point to the end byte of the block. hence
> > > the punch range is short, and this leads to unexpected behaviour in
> > > certain cases in xfs_bmap_punch_delalloc_range.
> > >
> > > e.g. pos = 0 means we call xfs_bmap_punch_delalloc_range(0,0), so
> > > there is no previous extent and it rounds up the punch to the end of
> > > the delalloc extent it found at offset 0, not the end of the range
> > > given to xfs_bmap_punch_delalloc_range().
> > >
> > > Fix this by handling the zero block offset case correctly.
> > >
> > > Fixes: 7348b322332d ("xfs: xfs_bmap_punch_delalloc_range() should take a byte range")
> > > Reported-by: Pengfei Xu <pengfei.xu@xxxxxxxxx>
> > > Found-by: Brian Foster <bfoster@xxxxxxxxxx>
> > > Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx>
> > > ---
> > > fs/xfs/xfs_aops.c | 14 ++++++++++++--
> > > 1 file changed, 12 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
> > > index 41734202796f..429f63cfd7d4 100644
> > > --- a/fs/xfs/xfs_aops.c
> > > +++ b/fs/xfs/xfs_aops.c
> > > @@ -466,6 +466,7 @@ xfs_discard_folio(
> > > {
> > > struct xfs_inode *ip = XFS_I(folio->mapping->host);
> > > struct xfs_mount *mp = ip->i_mount;
> > > + xfs_off_t end_off;
> > > int error;
> > >
> > > if (xfs_is_shutdown(mp))
> > > @@ -475,8 +476,17 @@ xfs_discard_folio(
> > > "page discard on page "PTR_FMT", inode 0x%llx, pos %llu.",
> > > folio, ip->i_ino, pos);
> > >
> > > - error = xfs_bmap_punch_delalloc_range(ip, pos,
> > > - round_up(pos, folio_size(folio)));
> > > + /*
> > > + * Need to be careful with the case where the pos passed in points to
> > > + * the first byte of the folio - rounding up won't change the value,
> > > + * but in all cases here we need to end offset to point to the start
> > > + * of the next folio.
> > > + */
> > > + if (pos == folio_pos(folio))
> > > + end_off = pos + folio_size(folio);
> > > + else
> > > + end_off = round_up(pos, folio_size(folio));
> >
> > Can this construct be simplified to:
> >
> > end_off = round_up(pos + 1, folio_size(folio));
>
> I thought about that first, but I really, really dislike sprinkling
> magic "+ 1" corrections into the code to address non-obvious
> unexplained off-by-one problems.
>
>
> > If pos is the first byte of the folio, it'll round end_off to the start
> > of the next folio. If pos is (somehow) the last byte of the folio, the
> > first argument to round_up is already the first byte of the next folio,
> > and rounding won't change it.
>
> Yup, and that's exactly the problem I had with doing this - it
> relies on the implicit behaviour that by moving last byte of a block
> to the first byte of the next block, round_up() won't change the end
> offset. i.e. the correct functioning of the code is just as
> non-obvious with a magic "+ 1" as the incorrect functioning was
> without it.
>
> Look at it this way: I didn't realise it was wrong when I wrote the
> code, and I couldn't find the bug round_up() introduced when reading
> the code even after the problem had been bisected to this exact
> change. The code I wrote is bad, and adding a magic "+ 1" to fix the
> bug doesn't make the code any better.
>
> Given this is a slow path, so I see no point in optimising the code
> for efficiency. IMO, clarity of the logic and calculation being made
> is far more important - obviously correct logic is better than
> relying on the effect of a magic "+ 1" on some other function to
> acheive the same thing....
<nod> Just making sure I wasn't missing something.
By the way, was this reported to the list?
Reviewed-by: Darrick J. Wong <djwong@xxxxxxxxxx>
--D
> Cheers,
>
> Dave.
> --
> Dave Chinner
> david@xxxxxxxxxxxxx
