[PATCH v3 2/4] drm: Add support for ARM's HDLCD controller.

[Liviu Dudau]

On Thu, Dec 03, 2015 at 02:20:26PM +0000, Daniel Stone wrote:
> Hi Liviu,
> 
> On 3 December 2015 at 10:00, Liviu Dudau <Liviu.Dudau at arm.com> wrote:
> > On Wed, Dec 02, 2015 at 05:21:44PM +0000, Daniel Stone wrote:
> >> On 2 December 2015 at 12:23, Liviu Dudau <Liviu.Dudau at arm.com> wrote:
> >> > +       if (irq_status & HDLCD_INTERRUPT_VSYNC) {
> >> > +               unsigned long flags;
> >> > +
> >> > +               drm_handle_vblank(drm, 0);
> >> > +
> >> > +               spin_lock_irqsave(&drm->event_lock, flags);
> >> > +               if (hdlcd->event) {
> >> > +                       drm_send_vblank_event(drm, hdlcd->event->pipe, hdlcd->event);
> >> > +                       drm_crtc_vblank_put(&hdlcd->crtc);
> >> > +                       hdlcd->event = NULL;
> >> > +               }
> >> > +               spin_unlock_irqrestore(&drm->event_lock, flags);
> >> > +       }
> >>
> >> As with VC4 and Rockchip, you're missing a ->preclose handler in your
> >> drm_drv, to make sure that you don't try to send events to a dead
> >> client (which causes an OOPS):
> >> https://git.collabora.com/cgit/user/daniels/linux.git/commit/?h=wip/4.4.x/rockchip-drm-fixes&id=d14f21bcd7
> >> (and its parent)
> >
> > Thanks for reviewing this!
> >
> > I do acknowledge your concerns and you might correct my understanding on
> > how atomic DRM works, but I believe in this case we should be safe. The
> > event stored in hdlcd->event is taken out of the crtc->state->event during
> > the crtc->atomic_begin() callback. If the client is dead the callback should
> > not be called, so that's how we avoid the OOPS.
> 
> Right, it's taken out of the CRTC state and put into the overall HDLCD
> structure. So the OOPS happens when:
>   - client submits state with event requested, async flag set
>   - atomic_begin moves crtc->state->event into hdlcd->event
>   - control returns to client, who exits immediately
>   - vblank happens
>   - hdlcd->event->base.file_priv now points to a dead client
>   - OOPS

I will try to construct an userspace test to see if I can trigger this scenario.

Thanks for explaining it to me.

Best regards,
Liviu

> 
> >> Also, is there anything preventing clients from submitting multiple
> >> pageflips before the event is sent? I couldn't see anything from a
> >> quick look, so you could have the situation of:
> >>   - client submits pageflip, event 1 stored to hdlcd->event
> >>   - client submits pageflip, event 2 stored to hdlcd->event
> >>   - vblank arrives, event 2 is sent
> >>   - event 1 has disappeared and been leaked
> >
> > As for multiple events being submitted before vsync interrupt happening: given
> > that the event is plucked out of the crtc->state, is that possible? And if so,
> > is there a case where having the later event overwrite the earlier one a desired
> > outcome?
> 
> Having events being overwritten is extremely undesirable; it could
> cause a client to stall in the right scenarios (e.g. requests
> submitted separately for different planes). It would be far better to
> turn hdlcd->event into a list of events which are sent per-vblank,
> probably just by retaining a list of pending states to apply; this
> also makes it easier to handle async commits in the future (which
> Weston in particular relies on).
> 
> Cheers,
> Daniel
> 

-- 
====================
| I would like to |
| fix the world,  |
| but they're not |
| giving me the   |
 \ source code!  /
  ---------------
    ¯\_(ツ)_/¯