Re: [RFC PATCH] mm: move xa forward when run across zombie page

[Zhaoyang Huang]

On Fri, Oct 14, 2022 at 8:12 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Fri, Oct 14, 2022 at 01:30:48PM +0800, zhaoyang.huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> >
> > Bellowing RCU stall is reported where kswapd traps in a live lock when shrink
> > superblock's inode list. The direct reason is zombie page keeps staying on the
> > xarray's slot and make the check and retry loop permanently. The root cause is unknown yet
> > and supposed could be an xa update without synchronize_rcu etc. I would like to
> > suggest skip this page to break the live lock as a workaround.
>
> No, the underlying bug should be fixed.
>
> >       if (!folio || xa_is_value(folio))
> >               return folio;
> >
> > -     if (!folio_try_get_rcu(folio))
> > +     if (!folio_try_get_rcu(folio)) {
> > +             xas_advance(xas, folio->index + folio_nr_pages(folio) - 1);
> >               goto reset;
> > +     }
>
> You can't do this anyway.  To call folio_nr_pages() and to look at
> folio->index, you must have a refcount on the page, and this is the
> path where we failed to get the refcount.
OK, could I move the xas like below?

+     if (!folio_try_get_rcu(folio)) {
+             xas_next_offset(xas);
              goto reset;
+     }