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

[Peter Xu]

On Wed, Oct 13, 2021 at 02:42:42PM -0700, Yang Shi wrote:
> On Tue, Oct 12, 2021 at 8:41 PM Peter Xu <peterx@xxxxxxxxxx> wrote:
> >
> > On Tue, Oct 12, 2021 at 08:27:06PM -0700, Yang Shi wrote:
> > > > But this also reminded me that shouldn't we be with the page lock already
> > > > during the process of "setting hwpoison-subpage bit, split thp, clear
> > > > hwpoison-subpage bit"?  If it's only the small window that needs protection,
> > > > while when looking up the shmem pagecache we always need to take the page lock
> > > > too, then it seems already safe even without the extra bit?  Hmm?
> > >
> > > I don't quite get your point. Do you mean memory_failure()? If so the
> > > answer is no, outside the page lock. And the window may be indefinite
> > > since file THP doesn't get split before this series and the split may
> > > fail even after this series.
> >
> > What I meant is that we could extend the page_lock in try_to_split_thp_page()
> > to cover setting hwpoison-subpage too (and it of course covers the clearing if
> > thp split succeeded, as that's part of the split process).  But yeah it's a
> > good point that the split may fail, so the extra bit seems still necessary.
> >
> > Maybe that'll be something worth mentioning in the commit message too?  The
> > commit message described very well on the overhead of looping over 512 pages,
> > however the reader can easily overlook the real reason for needing this bit -
> > IMHO it's really for the thp split failure case, as we could also mention that
> > if thp split won't fail, page lock should be suffice (imho).  We could also
> 
> Not only for THP split failure case. Before this series, shmem THP
> does't get split at all. And this commit is supposed to be backported
> to the older versions, so saying "page lock is sufficient" is not
> precise and confusing.

Sure, please feel free to use any wording you prefer as long as the other side
of the lock besides the performance impact could be mentioned.  Thanks,

-- 
Peter Xu