[PATCH 2/2] drm/atomic: clear new_state pointers at hw_done

[Rob Clark]

On Mon, Nov 4, 2019 at 12:50 PM Ville Syrjälä
<ville.syrjala at linux.intel.com> wrote:
>
> On Mon, Nov 04, 2019 at 11:13:59AM -0800, Rob Clark wrote:
> > On Mon, Nov 4, 2019 at 10:42 AM Ville Syrjälä
> > <ville.syrjala at linux.intel.com> wrote:
> > >
> > > On Fri, Nov 01, 2019 at 03:14:09PM -0700, Rob Clark wrote:
> > > > On Fri, Nov 1, 2019 at 2:44 PM Ville Syrjälä
> > > > <ville.syrjala at linux.intel.com> wrote:
> > > > >
> > > > > On Fri, Nov 01, 2019 at 12:49:02PM -0700, Rob Clark wrote:
> > > > > > On Fri, Nov 1, 2019 at 12:25 PM Ville Syrjälä
> > > > > > <ville.syrjala at linux.intel.com> wrote:
> > > > > > >
> > > > > > > On Fri, Nov 01, 2019 at 11:07:13AM -0700, Rob Clark wrote:
> > > > > > > > From: Rob Clark <robdclark at chromium.org>
> > > > > > > >
> > > > > > > > The new state should not be accessed after this point.  Clear the
> > > > > > > > pointers to make that explicit.
> > > > > > > >
> > > > > > > > This makes the error corrected in the previous patch more obvious.
> > > > > > > >
> > > > > > > > Signed-off-by: Rob Clark <robdclark at chromium.org>
> > > > > > > > ---
> > > > > > > >  drivers/gpu/drm/drm_atomic_helper.c | 29 +++++++++++++++++++++++++++++
> > > > > > > >  1 file changed, 29 insertions(+)
> > > > > > > >
> > > > > > > > diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
> > > > > > > > index 732bd0ce9241..176831df8163 100644
> > > > > > > > --- a/drivers/gpu/drm/drm_atomic_helper.c
> > > > > > > > +++ b/drivers/gpu/drm/drm_atomic_helper.c
> > > > > > > > @@ -2234,13 +2234,42 @@ EXPORT_SYMBOL(drm_atomic_helper_fake_vblank);
> > > > > > > >   */
> > > > > > > >  void drm_atomic_helper_commit_hw_done(struct drm_atomic_state *old_state)
> > > > > > > >  {
> > > > > > > > +     struct drm_connector *connector;
> > > > > > > > +     struct drm_connector_state *old_conn_state, *new_conn_state;
> > > > > > > >       struct drm_crtc *crtc;
> > > > > > > >       struct drm_crtc_state *old_crtc_state, *new_crtc_state;
> > > > > > > > +     struct drm_plane *plane;
> > > > > > > > +     struct drm_plane_state *old_plane_state, *new_plane_state;
> > > > > > > >       struct drm_crtc_commit *commit;
> > > > > > > > +     struct drm_private_obj *obj;
> > > > > > > > +     struct drm_private_state *old_obj_state, *new_obj_state;
> > > > > > > >       int i;
> > > > > > > >
> > > > > > > > +     /*
> > > > > > > > +      * After this point, drivers should not access the permanent modeset
> > > > > > > > +      * state, so we also clear the new_state pointers to make this
> > > > > > > > +      * restriction explicit.
> > > > > > > > +      *
> > > > > > > > +      * For the CRTC state, we do this in the same loop where we signal
> > > > > > > > +      * hw_done, since we still need to new_crtc_state to fish out the
> > > > > > > > +      * commit.
> > > > > > > > +      */
> > > > > > > > +
> > > > > > > > +     for_each_oldnew_connector_in_state(old_state, connector, old_conn_state, new_conn_state, i) {
> > > > > > > > +             old_state->connectors[i].new_state = NULL;
> > > > > > > > +     }
> > > > > > > > +
> > > > > > > > +     for_each_oldnew_plane_in_state(old_state, plane, old_plane_state, new_plane_state, i) {
> > > > > > > > +             old_state->planes[i].new_state = NULL;
> > > > > > > > +     }
> > > > > > > > +
> > > > > > > > +     for_each_oldnew_private_obj_in_state(old_state, obj, old_obj_state, new_obj_state, i) {
> > > > > > > > +             old_state->private_objs[i].new_state = NULL;
> > > > > > > > +     }
> > > > > > > > +
> > > > > > > >       for_each_oldnew_crtc_in_state(old_state, crtc, old_crtc_state, new_crtc_state, i) {
> > > > > > > >               old_state->crtcs[i].new_self_refresh_active = new_crtc_state->self_refresh_active;
> > > > > > > > +             old_state->crtcs[i].new_state = NULL;
> > > > > > >
> > > > > > > That's going to be a real PITA when doing programming after the fact from
> > > > > > > a vblank worker. It's already a pain that the new_crtc_state->state is
> > > > > > > getting NULLed somewhere.
> > > > > > >
> > > > > >
> > > > > > I think you already have that problem, this just makes it explicit.
> > > > >
> > > > > I don't yet. Except on a branch where I have my vblank workers.
> > > > > And I think the only problem is having the helpers/core clobber
> > > > > the pointers when it should not. I don't see why it can't just
> > > > > leave them be and let me use them.
> > > > >
> > > >
> > > > I guess it comes down to what assumptions you can make in driver
> > > > backend.  But if you can, for example, move planes between crtcs, I
> > > > think you can't make assumptions about the order in which things
> > > > complete even if you don't have commits overtaking each other on a
> > > > single crtc..
> > >
> > > IMO this whole notion of accessing new_crtc_state & co. being unsafe
> > > for some reason is wrong. I think as long as I have the drm_atomic_state
> > > I should be able to look at the new/old states within.
> > >
> >
> > accessing new state only works if you can guarantee the order in which
> > commits complete, which I don't think you can do in the general sense.
>
> Doesn't feel like it should take a lot of rocket science to guarantee
> the states get freed in the right order.
>

I agree, reference counting is not rocket science.  But that isn't the
way drm core handles per-object state currently.  Refcnt'ing is
probably the approach that I'd recommend if someone wanted to lift
this restriction about accessing new-state later.

BR,
-R