Re: [v3 PATCH 2/5] mm: filemap: check if THP has hwpoisoned subpage for PMD page fault

[Yang Shi]

On Thu, Oct 7, 2021 at 11:19 AM Yang Shi <shy828301@xxxxxxxxx> wrote:
>
> On Thu, Oct 7, 2021 at 9:06 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
> >
> > On Wed, Oct 06, 2021 at 04:57:38PM -0700, Yang Shi wrote:
> > > > For example, I see that both unpoison_memory() and soft_offline_page() will
> > > > call it too, does it mean that we'll also set the bits e.g. even when we want
> > > > to inject an unpoison event too?
> > >
> > > unpoison_memory() should be not a problem since it will just bail out
> > > once THP is met as the comment says:
> > >
> > > /*
> > > * unpoison_memory() can encounter thp only when the thp is being
> > > * worked by memory_failure() and the page lock is not held yet.
> > > * In such case, we yield to memory_failure() and make unpoison fail.
> > > */
> >
> > But I still think setting the subpage-hwpoison bit hides too deep there, it'll
> > be great we can keep get_hwpoison_page() as simple as a safe version of getting
> > the refcount of the page we want.  Or we'd still better touch up the comment
> > above get_hwpoison_page() to show that side effect.
> >
> > >
> > >
> > > And I think we should set the flag for soft offline too, right? The
> >
> > I'm not familiar with either memory failure or soft offline, so far it looks
> > right to me.  However..
> >
> > > soft offline does set the hwpoison flag for the corrupted sub page and
> > > doesn't split file THP,
> >
> > .. I believe this will become not true after your patch 5, right?
>
> But THP split may fail, right?
>
> >
> > > so it should be captured by page fault as well. And yes for poison injection.
> >
> > One more thing: besides thp split and page free, do we need to conditionally
> > drop the HasHwpoisoned bit when received an unpoison event?
>
> It seems not to me, as the above comment from unpoison_memory() says
> unpoison can encounter thp only when the thp is being worked by
> memory_failure() and the page lock is not held yet. So it just bails
> out.
>
> In addition, unpoison just works for software injected errors, not
> real hardware failure.
>
> >
> > If my understanding is correct, we may need to scan all the subpages there, to
> > make sure HasHwpoisoned bit reflects the latest status for the thp in question.
> >
> > >
> > > But your comment reminds me that get_hwpoison_page() is just called
> > > when !MF_COUNT_INCREASED, so it means MADV_HWPOISON still could
> > > escape. This needs to be covered too.
> >
> > Right, maybe that's also a clue that we shouldn't set the new page flag within
> > get_hwpoison_page(), since get_hwpoison_page() is actually well coupled with
> > MF_COUNT_INCREASED and all of them are only about refcounting of the pages.
>
> Yeah, maybe, as long as there is not early bail out in some error
> handling paths.

It seems fine to move setting the flag out of get_hwpoison_page() to
right before splitting THP so that both MF_COUNT_INCREASED and
!MF_COUNT_INCREASED could be covered.


>
> >
> > --
> > Peter Xu
> >