Re: [PATCH 3/5] shmem: shmem_writepage() split unlikely i915 THP

[Hugh Dickins]

On Tue, 1 Sep 2020, Yang Shi wrote:
> On Sun, Aug 30, 2020 at 2:04 PM Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> >
> > drivers/gpu/drm/i915/gem/i915_gem_shmem.c contains a shmem_writeback()
> > which calls shmem_writepage() from a shrinker: that usually works well
> > enough; but if /sys/kernel/mm/transparent_hugepage/shmem_enabled has
> > been set to "force" (documented as "Force the huge option on for all -
> > very useful for testing"), shmem_writepage() is surprised to be called
> > with a huge page, and crashes on the VM_BUG_ON_PAGE(PageCompound) (I
> > did not find out where the crash happens when CONFIG_DEBUG_VM is off).
> >
> > LRU page reclaim always splits the shmem huge page first: I'd prefer not
> > to demand that of i915, so check and split compound in shmem_writepage().
> >
> > Fixes: 2d6692e642e7 ("drm/i915: Start writeback from the shrinker")
> > Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> > Cc: stable@xxxxxxxxxxxxxxx # v5.3+
> > ---
> > I've marked this for stable just for the info, but the number of users
> > affected is very probably 1, so please feel free to delete that marking.
> >
> >  mm/shmem.c |    9 +++++++++
> >  1 file changed, 9 insertions(+)
> >
> > --- 5.9-rc2/mm/shmem.c  2020-08-16 17:32:50.693507198 -0700
> > +++ linux/mm/shmem.c    2020-08-28 17:35:08.326024349 -0700
> > @@ -1362,7 +1362,15 @@ static int shmem_writepage(struct page *
> >         swp_entry_t swap;
> >         pgoff_t index;
> >
> > -       VM_BUG_ON_PAGE(PageCompound(page), page);
> > +       /*
> > +        * If /sys/kernel/mm/transparent_hugepage/shmem_enabled is "force",
> > +        * then drivers/gpu/drm/i915/gem/i915_gem_shmem.c gets huge pages,
> > +        * and its shmem_writeback() needs them to be split when swapping.
> > +        */
> > +       if (PageTransCompound(page))
> > +               if (split_huge_page(page) < 0)
> > +                       goto redirty;
> 
> The change looks good to me. Acked-by: Yang Shi <shy828301@xxxxxxxxx>

Thanks.

> 
> Just a nit: it may be better to move the spilte after the !PageLocked
> assertion? Split needs page locked too.

I hadn't considered that, but I think it's best left as is:
split_huge_page_to_list() has its own 
	VM_BUG_ON_PAGE(!PageLocked(head), head);
to enforce its needs: think of the old BUG_ON(!PageLocked(page))
below as enforcing shmem's needs, checking that split_huge_page()
did not unlock it :)

> 
> > +
> >         BUG_ON(!PageLocked(page));
> >         mapping = page->mapping;
> >         index = page->index;
> >