Re: [syzbot] KASAN: use-after-free Read in copy_page_from_iter_atomic (2)

[Qu Wenruo]

On 2022/6/14 15:17, Christoph Hellwig wrote:
> On Mon, Jun 13, 2022 at 09:39:12PM +0200, David Sterba wrote:
> > On Fri, Jun 10, 2022 at 12:10:19AM -0700, syzbot wrote:
> > > syzbot has bisected this issue to:
> > >
> > > commit 4cd4aed63125ccd4efc35162627827491c2a7be7
> > > Author: Christoph Hellwig <hch@xxxxxx>
> > > Date:   Fri May 27 08:43:20 2022 +0000
> > >
> > >      btrfs: fold repair_io_failure into btrfs_repair_eb_io_failure
> >
> > Josef also reported a crash and found a bug in the patch, now added as
> > fixup that'll be in for-next:
>
> The patch looks correct to me.  Two things to note here:
>
>   - I hadn't realized you had queued up the series.  I've actually
>     started to merge some of my bio work with the bio split at
>     submission time work from Qu and after a few iterations I think
>     I would do the repair code a bit differently based on that.
>     Can you just drop the series for now?
>   - I find it interesting that syzbot hits btrfs metadata repair.
>     xfstests seems to have no coverage and I could not come up with
>     a good idea how to properly test it.  Does anyone have a good
>     idea on how to intentially corrupt metadata in a deterministic
>     way?

The same way as data?

map-logical to find the location of a mirror, write 4 bytes of zero into
the location, then call it a day.

Although for metadata, you may want to choose a metadata that would
definitely get read.
Thus tree root is a good candidate.

Thanks,
Qu