Re: [PATCH v15 6/7] soc: qcom: rpmh-rsc: Clear active mode configuration for wake TCS

[Doug Anderson]

Hi,

On Sun, Apr 5, 2020 at 10:08 PM Maulik Shah <mkshah@xxxxxxxxxxxxxx> wrote:
>
> Hi,
>
> On 4/3/2020 1:44 AM, Doug Anderson wrote:
> > Hi,
> >
> > On Tue, Mar 31, 2020 at 6:21 AM Maulik Shah <mkshah@xxxxxxxxxxxxxx> wrote:
> >> @@ -243,6 +279,14 @@ static irqreturn_t tcs_tx_done(int irq, void *p)
> >>                  }
> >>
> >>                  trace_rpmh_tx_done(drv, i, req, err);
> >> +
> >> +               /*
> >> +                * If wake tcs was re-purposed for sending active
> >> +                * votes, clear AMC trigger & enable modes and
> >> +                * disable interrupt for this TCS
> >> +                */
> >> +               if (!drv->tcs[ACTIVE_TCS].num_tcs)
> >> +                       __tcs_set_trigger(drv, i, false);
> > Still seems weird that we have to do the untrigger in the IRQ routine
> > here and also weird that we _don't_ do it in the IRQ routine for
> > non-borrowed TCSes.  I guess it's not the end of the world, though.
> >
> > Reviewed-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
>
> Thanks Doug for the review.
>
> IRQ is only needed to be enabled for TCSes used as ACTIVE_TCS.
>
> When we have dedicated ACTIVE_TCS, we leave IRQ always enabled from
> probe (one time configuration), since the TCS won't be used for anything
> other than to send ACTIVE transaction.
>
> When we don't have dedicated ACTIVE_TCS, we enable it when borrowed TCS
> is used for ACTIVE transaction and then once its done using it, we
> disable it again to leave it in its original configuration.

Sure, I get the concept.  ...but:

* It seems like it would be ideal to not call __tcs_set_trigger() from
the IRQ handler.  It uses write_tcs_reg_sync() which has a wait loop
in it.  That's not something you normally want in an IRQ handler.

* It seems like it would be a common case for the WAKE_TCS to be
borrowed several times in-between actually using it as a WAKE_TCS.

To make both of the above things better, it seems like you could just
leave the WAKE_TCS configured for active-mode transfers and just
switch all that stuff off when you actually need it as a WAKE_TCS.  To
some extent we could conceptually re-envision this and think of this
as being the ACTIVE_TCS that is occasionally borrowed to make up for
not having a WAKE_TCS.


In any case, I think the patch you have right now is good enough and
it feels like it's overdue to land this series.  I'd suggest against
spinning at this point.  Maybe we can try to improve this in a future
patch after your series lands.

-Doug