Re: [PATCH] iommufd/device: Enforce reserved IOVA also when attached to hwpt_nested

[Nicolin Chen <nicolinc@nvidia.com>]

On Tue, Aug 06, 2024 at 02:59:53AM +0000, Tian, Kevin wrote:
> > From: Nicolin Chen <nicolinc@nvidia.com>
> > Sent: Tuesday, August 6, 2024 2:04 AM
> >
> > On Mon, Aug 05, 2024 at 07:40:45AM +0000, Tian, Kevin wrote:
> > > > From: Nicolin Chen <nicolinc@nvidia.com>
> > > > Sent: Friday, August 2, 2024 1:35 PM
> > > >
> > > >  static int
> > > > -iommufd_group_do_replace_paging(struct iommufd_group *igroup,
> > > > -                             struct iommufd_hwpt_paging *hwpt_paging)
> > > > +iommufd_group_do_replace_reserved_iova(struct iommufd_group
> > *igroup,
> > > > +                                    struct iommufd_hw_pagetable *hwpt)
> > > >  {
> > > > +     struct iommufd_hwpt_paging *hwpt_paging =
> > > > to_hwpt_paging(hwpt);
> > > >       struct iommufd_hw_pagetable *old_hwpt = igroup->hwpt;
> > > >       struct iommufd_device *cur;
> > > >       int rc;
> > > >
> > > >       lockdep_assert_held(&igroup->lock);
> > > >
> > > > -     if (!hwpt_is_paging(old_hwpt) ||
> > > > -         hwpt_paging->ioas != to_hwpt_paging(old_hwpt)->ioas) {
> > > > +     if (!hwpt_paging)
> > > > +             return 0;
> > > > +
> > > > +     if (iommufd_hw_pagetable_compare_ioas(old_hwpt, hwpt)) {
> > >
> > > hmm this change is broken. In this helper:
> > >
> > >         if (!old_hwpt_paging || !new_hwpt_paging)
> > >                 return false;
> > >         return old_hwpt_paging->ioas != new_hwpt_paging->ioas;
> > >
> > > Obviously the original code wants to enforce reserved regions if
> > > new_hwpt is paging && old_hwpt is not paging, but this change
> > > skips this scenario.
> >
> > Hmm..I think that is the intention of this patch?
> >
> > The original code does that because it didn't enforce reserved
> > region (to the parent paging hwpt) when attaching the group to
> > a nested one. Now, it does. So, we basically check whether the
> > associated ioas has changed or not. Right?
> >
> 
> that is too much tied to nested vs. paging type. In case a new
> type comes w/o parent/ioas in the future, above will be broken.
> 
> IMHO to_hwpt_paging() already resolves the problem about
> nested parent while the original code above appears to be
> more future proof (w/ the enhanced to_hwpt_paging()), i.e.
> adding reserved region to the new hwpt w/ ioas as long as
> old/new hwpts don't belong to a same ioas (including the case
> where old has no ioas), instead of adding reserved region
> only when both old/new have an ioas but different. 😊

Ah, I see your point. It failed to cover the case where one of
the hwpts returns NULL from to_hwpt_paging().

Will fix and respin.

Thanks
Nicolin