RE: [PATCH v1 7/8] iommufd/device: Use iommu_group_replace_domain()

[Tian, Kevin]

> From: Nicolin Chen <nicolinc@xxxxxxxxxx>
> Sent: Tuesday, February 7, 2023 3:18 AM
> 
> On Mon, Feb 06, 2023 at 08:46:04AM +0000, Tian, Kevin wrote:
> > > From: Nicolin Chen <nicolinc@xxxxxxxxxx>
> > > Sent: Thursday, February 2, 2023 3:05 PM
> > >
> > > @@ -246,6 +249,18 @@ static int iommufd_device_do_attach(struct
> > > iommufd_device *idev,
> > >               }
> > >       }
> > >
> > > +     if (cur_hwpt) {
> > > +             /* Replace the cur_hwpt */
> > > +             mutex_lock(&cur_hwpt->devices_lock);
> > > +             if (cur_hwpt->ioas != hwpt->ioas)
> > > +                     iopt_remove_reserved_iova(&cur_hwpt->ioas->iopt,
> > > +                                               idev->dev);
> > > +             list_del(&cur_hwpt->hwpt_item);
> >
> > emmm shouldn't this be done only when the device is the last
> > one attached to the hwpt? and if it's the last one you should
> > also iopt_table_remove_domain() together with list_del, i.e.
> > similar housekeeping as done in iommufd_device_detach().
> 
> You are right. I had another patch on top of this series,
> moving this list_del() and iopt_table_remove_domain() to
> the destroy() callback, so I overlooked.
> 
> And I just found that the list_add_del(hwpt_item) in the
> IOMMUFD_OBJ_HW_PAGETABLE case doesn't seem to call at the
> first device's attachment. So, I think that we might need
> my previous "symmetric" patch in this series too.
> 

Yes, that makes sense.