Re: hwpoison, shmem: fix data lost issue for 5.15.y

[Naoya Horiguchi <naoya.horiguchi@linux.dev>]

On Wed, Nov 23, 2022 at 10:54:15AM -0800, Yang Shi wrote:
> On Tue, Nov 22, 2022 at 5:05 PM Mike Kravetz <mike.kravetz@oracle.com> wrote:
> >
> > On 11/15/22 01:16, HORIGUCHI NAOYA(堀口 直也) wrote:
> > > On Mon, Nov 14, 2022 at 02:53:51PM -0800, Mike Kravetz wrote:
> > > > On 11/15/22 07:39, Naoya Horiguchi wrote:
> > > > > On Mon, Nov 14, 2022 at 05:11:35PM +0100, Greg KH wrote:
> > > > > > On Mon, Nov 14, 2022 at 10:14:03PM +0900, Naoya Horiguchi wrote:
> > > > > > > Hi,
> > > > > > >
> > > > > > > I'd like to request the follow commits to be backported to 5.15.y.
> > > > > > >
> > > > > > > - dd0f230a0a80 ("mm: hwpoison: refactor refcount check handling")
> > > > > > > - 4966455d9100 ("mm: hwpoison: handle non-anonymous THP correctly")
> > > > > > > - a76054266661 ("mm: shmem: don't truncate page if memory failure happens")
> > > > > > >
> > > > > > > These patches fixed a data lost issue by preventing shmem pagecache from
> > > > > > > being removed by memory error.  These were not tagged for stable originally,
> > > > > > > but that's revisited recently.
> > > > > >
> > > > > > And have you tested that these all apply properly (and in which order?)
> > > > >
> > > > > Yes, I've checked that these cleanly apply (without any change) on
> > > > > 5.15.78 in the above order (i.e. dd0f23 is first, 496645 comes next,
> > > > > then a76054).
> > > > >
> > > > > > and work correctly?
> > > > >
> > > > > Yes, I ran related testcases in my test suite, and their status changed
> > > > > FAIL to PASS with these patches.
> > > >
> > > > Hi Naoya,
> > > >
> > > > Just curious if you have plans to do backports for earlier releases?
> > >
> > > I didn't have a clear plan.  I just thought that we should backport to
> > > earlier kernels if someone want and the patches are applicable easily
> > > enough and well-tested.
> > >
> > > >
> > > > If not, I can start that effort.  We have seen data loss/corruption because of
> > > > this on a 4.14 based release.   So, I would go at least that far back.
> > >
> > > Thank you for raising hand, that's really helpful.
> > >
> > > Maybe dd0f230a0a80 ("[PATCH] hugetlbfs: don't delete error page from
> > > pagecbache") should be considered to backport together, because it's
> > > the similar issue and reported (a while ago) to fail to backport.
> > > dd0f230a0a80 does not apply cleanly on top of 5.15.78 + the above 3 patches.
> > > So I need check more and will update my current proposal for 5.15.y.
> >
> > When working with 5.10.y, I noticed that commit eac96c3efdb5 ("mm: filemap:
> > check if THP has hwpoisoned subpage for PMD page fault") as well as the
> > prereq commit c7cb42e94473 ("mm: hwpoison: remove the unnecessary THP check")
> > were not backported to 5.10.y.  Without those patches, THP testing will
> > fail.
> >
> > Naoya and Yang Shi, does that sound right?
>
> Yes, since the hwpoisoned THP will be kept in page cache so the page
> fault may happen on it again, without that commit the page fault won't
> return -EHWPOISON if I remember correctly.
>
> >
> > I have backports for those as well but want to check if you think
> > anything else is needed.
>
> Thanks for backporting them. No more fix is needed AFAICT.

I agree with Yang.  There seems no other commit related to current
pagecache problem but not backported yet.

Thanks,
Naoya Horiguchi