Re: [PATCH] xfs: flush new eof page on truncate to avoid post-eof corruption

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Wed, Oct 07, 2020 at 10:35:09AM -0400, Brian Foster wrote:
> It is possible to expose non-zeroed post-EOF data in XFS if the new
> EOF page is dirty, backed by an unwritten block and the truncate
> happens to race with writeback. iomap_truncate_page() will not zero
> the post-EOF portion of the page if the underlying block is
> unwritten. The subsequent call to truncate_setsize() will, but
> doesn't dirty the page. Therefore, if writeback happens to complete
> after iomap_truncate_page() (so it still sees the unwritten block)
> but before truncate_setsize(), the cached page becomes inconsistent
> with the on-disk block. A mapped read after the associated page is
> reclaimed or invalidated exposes non-zero post-EOF data.
> 
> For example, consider the following sequence when run on a kernel
> modified to explicitly flush the new EOF page within the race
> window:
> 
> $ xfs_io -fc "falloc 0 4k" -c fsync /mnt/file
> $ xfs_io -c "pwrite 0 4k" -c "truncate 1k" /mnt/file
>   ...
> $ xfs_io -c "mmap 0 4k" -c "mread -v 1k 8" /mnt/file
> 00000400:  00 00 00 00 00 00 00 00  ........
> $ umount /mnt/; mount <dev> /mnt/
> $ xfs_io -c "mmap 0 4k" -c "mread -v 1k 8" /mnt/file
> 00000400:  cd cd cd cd cd cd cd cd  ........
> 
> Update xfs_setattr_size() to explicitly flush the new EOF page prior
> to the page truncate to ensure iomap has the latest state of the
> underlying block.
> 
> Fixes: 68a9f5e7007c ("xfs: implement iomap based buffered write path")
> Signed-off-by: Brian Foster <bfoster@redhat.com>
> ---
> 
> This patch is intentionally simplistic because I wanted to get some
> thoughts on a proper fix and at the same time consider something easily
> backportable. The iomap behavior seems rather odd to me in general,
> particularly if we consider the same kind of behavior can occur on
> file-extending writes. It's just not a user observable problem in that
> case because a sub-page write of a current EOF page (backed by an
> unwritten block) will zero fill the rest of the page at write time
> (before the zero range essentially skips it due to the unwritten block).
> It's not totally clear to me if that's an intentional design
> characteristic of iomap or something we should address.
> 
> It _seems_ like the more appropriate fix is that iomap truncate page
> should at least accommodate a dirty page over an unwritten block and
> modify the page (or perhaps just unconditionally do a buffered write on
> a non-aligned truncate, similar to what block_truncate_page() does). For
> example, we could push the UNWRITTEN check from iomap_zero_range_actor()
> down into iomap_zero(), actually check for an existing page there, and
> then either zero it or skip out if none exists. Thoughts?

I haven't looked at this in much depth yet, but I agree with the
principle that iomap ought to handle the case of unwritten extents
fronted by dirty pagecache.

--D

> Brian
> 
>  fs/xfs/xfs_iops.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/fs/xfs/xfs_iops.c b/fs/xfs/xfs_iops.c
> index 80a13c8561d8..3ef2e77b454e 100644
> --- a/fs/xfs/xfs_iops.c
> +++ b/fs/xfs/xfs_iops.c
> @@ -911,6 +911,16 @@ xfs_setattr_size(
>  		error = iomap_zero_range(inode, oldsize, newsize - oldsize,
>  				&did_zeroing, &xfs_buffered_write_iomap_ops);
>  	} else {
> +		/*
> +		 * iomap won't detect a dirty page over an unwritten block and
> +		 * subsequently skips zeroing the newly post-eof portion of the
> +		 * page. Flush the new EOF to convert the block before the
> +		 * pagecache truncate.
> +		 */
> +		error = filemap_write_and_wait_range(inode->i_mapping, newsize,
> +						     newsize);
> +		if (error)
> +			return error;
>  		error = iomap_truncate_page(inode, newsize, &did_zeroing,
>  				&xfs_buffered_write_iomap_ops);
>  	}
> -- 
> 2.25.4
>