Re: [PATCH v1] drm/msm/dp: use dp_hpd_plug_handle() and dp_hpd_unplug_handle() directly

[Dmitry Baryshkov]

On Fri, 29 Mar 2024 at 04:16, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
>
>
>
> On 3/28/2024 5:10 PM, Dmitry Baryshkov wrote:
> > On Fri, 29 Mar 2024 at 01:42, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
> >>
> >>
> >>
> >> On 3/28/2024 3:50 PM, Dmitry Baryshkov wrote:
> >>> On Thu, 28 Mar 2024 at 23:21, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
> >>>>
> >>>>
> >>>>
> >>>> On 3/28/2024 1:58 PM, Stephen Boyd wrote:
> >>>>> Quoting Abhinav Kumar (2024-03-28 13:24:34)
> >>>>>> + Johan and Bjorn for FYI
> >>>>>>
> >>>>>> On 3/28/2024 1:04 PM, Kuogee Hsieh wrote:
> >>>>>>> For internal HPD case, hpd_event_thread is created to handle HPD
> >>>>>>> interrupts generated by HPD block of DP controller. It converts
> >>>>>>> HPD interrupts into events and executed them under hpd_event_thread
> >>>>>>> context. For external HPD case, HPD events is delivered by way of
> >>>>>>> dp_bridge_hpd_notify() under thread context. Since they are executed
> >>>>>>> under thread context already, there is no reason to hand over those
> >>>>>>> events to hpd_event_thread. Hence dp_hpd_plug_handle() and
> >>>>>>> dp_hpd_unplug_hanlde() are called directly at dp_bridge_hpd_notify().
> >>>>>>>
> >>>>>>> Signed-off-by: Kuogee Hsieh <quic_khsieh@xxxxxxxxxxx>
> >>>>>>> ---
> >>>>>>>      drivers/gpu/drm/msm/dp/dp_display.c | 5 +++--
> >>>>>>>      1 file changed, 3 insertions(+), 2 deletions(-)
> >>>>>>>
> >>>>>>
> >>>>>> Fixes: 542b37efc20e ("drm/msm/dp: Implement hpd_notify()")
> >>>>>
> >>>>> Is this a bug fix or an optimization? The commit text doesn't tell me.
> >>>>>
> >>>>
> >>>> I would say both.
> >>>>
> >>>> optimization as it avoids the need to go through the hpd_event thread
> >>>> processing.
> >>>>
> >>>> bug fix because once you go through the hpd event thread processing it
> >>>> exposes and often breaks the already fragile hpd handling state machine
> >>>> which can be avoided in this case.
> >>>
> >>> Please add a description for the particular issue that was observed
> >>> and how it is fixed by the patch.
> >>>
> >>> Otherwise consider there to be an implicit NAK for all HPD-related
> >>> patches unless it is a series that moves link training to the enable
> >>> path and drops the HPD state machine completely.
> >>>
> >>> I really mean it. We should stop beating a dead horse unless there is
> >>> a grave bug that must be fixed.
> >>>
> >>
> >> I think the commit message is explaining the issue well enough.
> >>
> >> This was not fixing any issue we saw to explain you the exact scenario
> >> of things which happened but this is just from code walkthrough.
> >>
> >> Like kuogee wrote, hpd event thread was there so handle events coming
> >> out of the hpd_isr for internal hpd cases. For the hpd_notify coming
> >> from pmic_glink or any other extnernal hpd cases, there is no need to
> >> put this through the hpd event thread because this will only make things
> >> worse of exposing the race conditions of the state machine.
> >>
> >> Moving link training to enable and removal of hpd event thread will be
> >> worked on but delaying obvious things we can fix does not make sense.
> >
> >  From the commit message this feels like an optimisation rather than a
> > fix. And granted the fragility of the HPD state machine, I'd prefer to
> > stay away from optimisations. As far as I understood from the history
> > of the last revert, we'd better make sure that HPD handling goes only
> > through the HPD event thread.
> >
>
> I think you are mixing the two. We tried to send the events through
> DRM's hpd_notify which ended up in a bad way and btw, thats still not
> resolved even though I have seen reports that things are fine with the
> revert, we are consistently able to see us ending up in a disconnected
> state with all the reverts and fixes in our x1e80100 DP setup.
>
> I plan to investigate that issue properly in the next week and try to
> make some sense of it all.
>
> In fact, this patch is removing one more user of the hpd event thread
> which is the direction in which we all want to head towards.

As I stated earlier, from my point of view it doesn't make sense to
rework the HPD thread in small steps.

> On whether this is an optimization or a bug fix. I think by avoiding hpd
> event thread (which should have never been used for hpd_notify updates,
> hence a bug) we are avoiding the possibility of more race conditions.

I think that the HPD event thread serializes handling of events, so
avoiding it increases the possibility of a race condition.

>
> So, this has my R-b and it holds. Upto you.

I'd wait for a proper description of the issue that was observed and
how it is solved by this patch.

-- 
With best wishes
Dmitry