On Mon, May 20, 2019 at 09:23:00PM +0200, Daniel Vetter wrote:
> On Mon, May 20, 2019 at 10:09:24PM +0300, Imre Deak wrote:
> > On Mon, May 20, 2019 at 08:37:46PM +0200, Daniel Vetter wrote:
> > > On Mon, May 20, 2019 at 08:41:09PM +0300, Imre Deak wrote:
> > > > We allowed modesetting an unregistered connector only in the case the
> > > > mode is getting disabled on the connector.
> > > >
> > > > The reason for this check was the lack of proper refcounting for the
> > > > backing memory objects. That problem has been solved meanwhile so there
> > > > is no reason any more to reject the modesetting in general.
> > >
> > > I'm not parsing this at all ... maybe references to the commits that fix
> > > this? Or do you mean the refcounting work for all the things hanging of
> > > connectors, including the entire mst tree?
> >
> > Yes the check was added to solve the issue related to the removal of MST
> > connectors that could happen asynchronously wrt. a modeset referring to
> > that MST connector. That could happen since the MST core doesn't hold
> > any locks (for instance the connection_mutex) during removing an MST
> > connector that would prevent doing a modeset at the same time.
> >
> > Adding the refcounting for such MST connectors (via the
> > drm_connector_get()/drm_connector_put()) got rid of the above problem.
>
> We added the check way after that stuff landed. Before all the connector
> reworking connectors were forcefully disabled by the kernel. The idea
> behind this check is to make sure that that userspace notices a connector
> is gone (only thing that's not allowed is enabling it, you can keep
> pageflipping). I think we've always had behaviour like ever since mst (all
> userspace has some "oops mst connector probably gone" failure catching
> around modesets).
Right, pageflipping works.
> So no idea what all blows up if we stop catching userspace this way.
>
> Now very much possible I'm getting all this wrong again or missing
> something, this stuff is often way over my head. But I'm really vary of
> breaking userspace here. E.g. just the drm_connector_get/put lifetime
> changes results in some userspace breaking if you unplug/replug fast
> enough, because then it doesn't notice the connector change anymore. I
> couldn't figure out a way to paper over that regression without
> reintroduce the rather broken and oops-prone old connector lifetime
> management.
Yes, but what is the the actual description of the failing scenario? I
can't see how anything can go wrong without this check. The resume time
restoration modeset may have to act in the same way on an old connector.
I don't understand how userspace would not notice the connector change.
It will get a hotplug uevent in response to which it would have to do a
detect which returns to it the updated information about the new MST
connector tree.
> > > > The check
> > > > for that also makes driver internal modesets more cumbersome where we
> > > > need to add exemptions for the cases where we do need to allow the
> > > > modeset even for unregistered connectors. One such case is the
> > > > restoration of the mode during resume.
> > >
> > > Yeah this one actually makes sense to me. We could still keep this check
> > > here, but for the atomic ioctl only when called from userspace. But iirc
> > > Lyude also said she has some plans here, so no idea whether that all fits.
> > >
> > > > Simplify things by removing the unneeded check. I can't see how
> > > > modesetting an unregistered connector can cause any problem and the race
> > > > (described in the code comment) can anyway result in such a modeset (if
> > > > the connector is unregistered right after the check).
> > >
> > > Not saying we don't need this, but there's fairly enormous amounts of
> > > history behind all this stuff, and lots of discussions. Would be good to
> > > at least reference those, so we have a good story for when this then all
> > > goes wrong again.
> >
> > I still don't see why this check is needed. There is no justification
> > for it - besides the original reason for it as discussed above about the
> > refcounting problem, which is solved now - so I think we should remove
> > it, instead of just making it a special case for the user space modeset.
> >
> > As I wrote a user space modeset can end up anyway doing a modeset on an
> > unregistered connector when the unregistering - by MST core - happens just
> > right after the check.
>
> Yup. Always been like that.
> -Daniel
>
> >
> > > -Daniel
> > >
> > > >
> > > > Cc: Lyude Paul <lyude@xxxxxxxxxx>
> > > > Cc: Daniel Vetter <daniel.vetter@xxxxxxxx>
> > > > Cc: Ville Syrjälä <ville.syrjala@xxxxxxxxxxxxxxx>
> > > > Signed-off-by: Imre Deak <imre.deak@xxxxxxxxx>
> > > > ---
> > > > drivers/gpu/drm/drm_atomic_helper.c | 29 ++---------------------------
> > > > 1 file changed, 2 insertions(+), 27 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
> > > > index 2e0cb4246cbd..e94e69483498 100644
> > > > --- a/drivers/gpu/drm/drm_atomic_helper.c
> > > > +++ b/drivers/gpu/drm/drm_atomic_helper.c
> > > > @@ -319,33 +319,6 @@ update_connector_routing(struct drm_atomic_state *state,
> > > > return 0;
> > > > }
> > > >
> > > > - crtc_state = drm_atomic_get_new_crtc_state(state,
> > > > - new_connector_state->crtc);
> > > > - /*
> > > > - * For compatibility with legacy users, we want to make sure that
> > > > - * we allow DPMS On->Off modesets on unregistered connectors. Modesets
> > > > - * which would result in anything else must be considered invalid, to
> > > > - * avoid turning on new displays on dead connectors.
> > > > - *
> > > > - * Since the connector can be unregistered at any point during an
> > > > - * atomic check or commit, this is racy. But that's OK: all we care
> > > > - * about is ensuring that userspace can't do anything but shut off the
> > > > - * display on a connector that was destroyed after it's been notified,
> > > > - * not before.
> > > > - *
> > > > - * Additionally, we also want to ignore connector registration when
> > > > - * we're trying to restore an atomic state during system resume since
> > > > - * there's a chance the connector may have been destroyed during the
> > > > - * process, but it's better to ignore that then cause
> > > > - * drm_atomic_helper_resume() to fail.
> > > > - */
> > > > - if (!state->duplicated && drm_connector_is_unregistered(connector) &&
> > > > - crtc_state->active) {
> > > > - DRM_DEBUG_ATOMIC("[CONNECTOR:%d:%s] is not registered\n",
> > > > - connector->base.id, connector->name);
> > > > - return -EINVAL;
> > > > - }
> > > > -
> > > > funcs = connector->helper_private;
> > > >
> > > > if (funcs->atomic_best_encoder)
> > > > @@ -390,6 +363,8 @@ update_connector_routing(struct drm_atomic_state *state,
> > > >
> > > > set_best_encoder(state, new_connector_state, new_encoder);
> > > >
> > > > + crtc_state = drm_atomic_get_new_crtc_state(state,
> > > > + new_connector_state->crtc);
> > > > crtc_state->connectors_changed = true;
> > > >
> > > > DRM_DEBUG_ATOMIC("[CONNECTOR:%d:%s] using [ENCODER:%d:%s] on [CRTC:%d:%s]\n",
> > > > --
> > > > 2.17.1
> > > >
> > >
> > > --
> > > Daniel Vetter
> > > Software Engineer, Intel Corporation
> > > http://blog.ffwll.ch
>
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
