Messages in this thread | | | Date | Thu, 28 Mar 2024 22:47:03 -0700 | Subject | Re: [PATCH v1] drm/msm/dp: use dp_hpd_plug_handle() and dp_hpd_unplug_handle() directly | From | Abhinav Kumar <> |
| |
On 3/28/2024 8:23 PM, Dmitry Baryshkov wrote: > On Fri, 29 Mar 2024 at 04:16, Abhinav Kumar <quic_abhinavk@quicinc.com> wrote: >> >> >> >> On 3/28/2024 5:10 PM, Dmitry Baryshkov wrote: >>> On Fri, 29 Mar 2024 at 01:42, Abhinav Kumar <quic_abhinavk@quicinc.com> wrote: >>>> >>>> >>>> >>>> On 3/28/2024 3:50 PM, Dmitry Baryshkov wrote: >>>>> On Thu, 28 Mar 2024 at 23:21, Abhinav Kumar <quic_abhinavk@quicinc.com> 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@quicinc.com> >>>>>>>>> --- >>>>>>>>> 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. >
This was a code walkthrough fix as I wrote a few times. If there no merit in pushing this, lets ignore it and stop discussing.
| |