Re: [PATCH RESEND v6 1/9] pagemap: Introduce ->memory_failure()

[Dan Williams <dan.j.williams@intel.com>]

On Wed, Aug 18, 2021 at 12:52 AM ruansy.fnst@fujitsu.com
<ruansy.fnst@fujitsu.com> wrote:
>
>
>
> > -----Original Message-----
> > From: Jane Chu <jane.chu@oracle.com>
> > Subject: Re: [PATCH RESEND v6 1/9] pagemap: Introduce ->memory_failure()
> >
> >
> > On 8/17/2021 10:43 PM, Jane Chu wrote:
> > > More information -
> > >
> > > On 8/16/2021 10:20 AM, Jane Chu wrote:
> > >> Hi, ShiYang,
> > >>
> > >> So I applied the v6 patch series to my 5.14-rc3 as it's what you
> > >> indicated is what v6 was based at, and injected a hardware poison.
> > >>
> > >> I'm seeing the same problem that was reported a while ago after the
> > >> poison was consumed - in the SIGBUS payload, the si_addr is missing:
> > >>
> > >> ** SIGBUS(7): canjmp=1, whichstep=0, **
> > >> ** si_addr(0x(nil)), si_lsb(0xC), si_code(0x4, BUS_MCEERR_AR) **
> > >>
> > >> The si_addr ought to be 0x7f6568000000 - the vaddr of the first page
> > >> in this case.
> > >
> > > The failure came from here :
> > >
> > > [PATCH RESEND v6 6/9] xfs: Implement ->notify_failure() for XFS
> > >
> > > +static int
> > > +xfs_dax_notify_failure(
> > > ...
> > > +    if (!xfs_sb_version_hasrmapbt(&mp->m_sb)) {
> > > +        xfs_warn(mp, "notify_failure() needs rmapbt enabled!");
> > > +        return -EOPNOTSUPP;
> > > +    }
> > >
> > > I am not familiar with XFS, but I have a few questions I hope to get
> > > answers -
> > >
> > > 1) What does it take and cost to make
> > >     xfs_sb_version_hasrmapbt(&mp->m_sb) to return true?
>
> Enable rmpabt feature when making xfs filesystem
>    `mkfs.xfs -m rmapbt=1 /path/to/device`
> BTW, reflink is enabled by default.
>
> > >
> > > 2) For a running environment that fails the above check, is it
> > >     okay to leave the poison handle in limbo and why?
> It will fall back to the old handler.  I think you have already known it.
>
> > >
> > > 3) If the above regression is not acceptable, any potential remedy?
> >
> > How about moving the check to prior to the notifier registration?
> > And register only if the check is passed?  This seems better than an
> > alternative which is to fall back to the legacy memory_failure handling in case
> > the filesystem returns -EOPNOTSUPP.
>
> Sounds like a nice solution.  I think I can add an is_notify_supported() interface in dax_holder_ops and check it when register dax_holder.

Shouldn't the fs avoid registering a memory failure handler if it is
not prepared to take over? For example, shouldn't this case behave
identically to ext4 that will not even register a callback?