Re: [PATCH v4 29/78] drm/vc4: crtc: Add a delay after disabling the PixelValve output

[Maxime Ripard]

Hi Stefan

On Tue, Aug 25, 2020 at 11:30:58PM +0200, Stefan Wahren wrote:
> Am 25.08.20 um 17:06 schrieb Maxime Ripard:
> > Hi Stefan,
> >
> > On Wed, Jul 29, 2020 at 05:50:31PM +0200, Stefan Wahren wrote:
> >> Am 29.07.20 um 16:42 schrieb Maxime Ripard:
> >>> Hi,
> >>>
> >>> On Wed, Jul 29, 2020 at 03:09:21PM +0100, Dave Stevenson wrote:
> >>>> On Wed, 8 Jul 2020 at 18:43, Maxime Ripard <maxime@xxxxxxxxxx> wrote:
> >>>>> In order to avoid pixels getting stuck in the (unflushable) FIFO between
> >>>>> the HVS and the PV, we need to add some delay after disabling the PV output
> >>>>> and before disabling the HDMI controller. 20ms seems to be good enough so
> >>>>> let's use that.
> >>>>>
> >>>>> Signed-off-by: Maxime Ripard <maxime@xxxxxxxxxx>
> >>>>> ---
> >>>>>  drivers/gpu/drm/vc4/vc4_crtc.c | 2 ++
> >>>>>  1 file changed, 2 insertions(+)
> >>>>>
> >>>>> diff --git a/drivers/gpu/drm/vc4/vc4_crtc.c b/drivers/gpu/drm/vc4/vc4_crtc.c
> >>>>> index d0b326e1df0a..7b178d67187f 100644
> >>>>> --- a/drivers/gpu/drm/vc4/vc4_crtc.c
> >>>>> +++ b/drivers/gpu/drm/vc4/vc4_crtc.c
> >>>>> @@ -403,6 +403,8 @@ static void vc4_crtc_atomic_disable(struct drm_crtc *crtc,
> >>>>>         ret = wait_for(!(CRTC_READ(PV_V_CONTROL) & PV_VCONTROL_VIDEN), 1);
> >>>>>         WARN_ONCE(ret, "Timeout waiting for !PV_VCONTROL_VIDEN\n");
> >>>>>
> >>>>> +       mdelay(20);
> >>>> mdelay for 20ms seems a touch unfriendly as it's a busy wait. Can we
> >>>> not msleep instead?
> >>> Since the timing was fairly critical, sleeping didn't seem like a good
> >>> solution since there's definitely some chance you overshoot and end up
> >>> with a higher time than the one you targeted.
> >> usleep_range(min, max) isn't a solution?
> > My understanding of usleep_range was that you can still overshoot, even
> > though it's backed by an HR timer so the resolution is not a jiffy. Are
> > we certain that we're going to be in that range?
> 
> you are right there is no guarantee about the upper wake up time.
> 
> And it's not worth the effort to poll the FIFO state until its empty
> (using 20 ms as timeout)?

I know this isn't really a great argument there, but getting this to
work has been quite painful, and the timing is very sensitive. If we
fail to wait for enough time, there's going to be a pixel shift that we
can't get rid of unless we reboot, which is pretty bad (and would fail
any CI test that checks for the output integrity).

I know busy-looping for 20ms isn't ideal, but it's not really in a
hot-path (it's only done when changing a mode), with the sync time of
the display likely to be much more than that, and if it can avoid having
to look into it ever again or avoid random failures, I'd say it's worth
it.

Maxime

Attachment: signature.asc
Description: PGP signature