On Mon, 10 May 2021 at 08:37, Rajendra Nayak <rnayak@xxxxxxxxxxxxxx> wrote:
>
>
>
> On 5/7/2021 2:36 PM, Ulf Hansson wrote:
> > On Tue, 4 May 2021 at 09:18, Rajendra Nayak <rnayak@xxxxxxxxxxxxxx> wrote:
> >>
> >>
> >> []...
> >>>>>>
> >>>>>> Ulf, Viresh, I think we discussed this at the time of introducing the
> >>>>>> performance states.
> >>>>>>
> >>>>>> The client's state does not affect if its performance_state should
> >>>>>> be included in the calculation of the aggregated performance_state, so
> >>>>>> each driver that needs to keep some minimum performance state needs to
> >>>>>> have these two snippets.
> >>>>>>
> >>>>>> Would it not make sense to on enable/disable re-evaluate the
> >>>>>> performance_state and potentially reconfigure the hardware
> >>>>>> automatically?
> >>>>>
> >>>>> I agree, this will be repeated across multiple drivers which would
> >>>>> need some minimal vote while they are active, handling this during
> >>>>> genpd enable/disable in genpd core makes sense.
> >>>>
> >>>> Initially that's what we tried out, but we realized that it was
> >>>> difficult to deal with this internally in genpd, but more importantly
> >>>> it also removed some flexibility from consumers and providers. See
> >>>> commit 68de2fe57a8f ("PM / Domains: Make genpd performance states
> >>>> orthogonal to the idlestates").
> >>>>
> >>>> As a matter of fact this was quite recently discussed [1], which also
> >>>> pointed out some issues when using the "required-opps" in combination,
> >>>> but perhaps that got resolved? Viresh?
> >>>
> >>> So I looked again at that thread in detail today. The basic idea was
> >>> to enable/disable the genpd from within the OPP core and there were
> >>> doubts on how to do that efficiently as there are cases where domains
> >>> may be enabled for an OPP, but not for others.. etc. etc.
> >>>
> >>> I am not sure if I consider that thread as part of the discussion we
> >>> are having here, they may be related, but that thread doesn't block
> >>> anything to be done in the genpd core.
> >>
> >> That's true, the 2 threads are different in the sense that one talks
> >> about having OPP core managing power on/off along with setting perf state,
> >> while the other talks about genpd core managing a default perf state
> >> along with power on/off, but they are similar in the sense that both
> >> are related to the discussion whether we should treat powering on and off
> >> a domain related to setting its performance state or if it should be
> >> considered completely orthogonal.
> >>
> >> I think the clock framework treats setting clock rates and turning
> >> on/off a clock orthogonal because there is an inherent assumption that
> >> once the clock is turned off, what rate it was set to should not matter,
> >> and it can be running at the same rate when we turn the clock back on.
> >>
> >> I guess we can have the same assumption here that a perf state of a
> >> power domain should not matter if the power domain is turned off
> >> and hence the perf state need not be dropped explicitly during power off,
> >> atleast that should be true for the qcom power domains supporting perf
> >> state upstream.
> >>
> >> Should that be the approach taken here? I guess that would mean the patch
> >> I had proposed earlier [1] to manage this in the genpd core would have to set the default
> >> perf state at attach and remove it only during a detach of the device to
> >> the pm_domain, and not manage it during the runtime_suspend/resume of the device.
> >
> > Right, I think this would be a step in the right direction, but it's
> > not sufficient to solve the complete problem. As you also point out
> > below.
> >
> >>
> >>>> A consumer driver
> >>>> can no longer make its vote for its device to stick around, when the
> >>>> device becomes runtime suspended - and how do we know that we never
> >>>> need to support such a case?
> >>
> >> The above approach should take care of this but the down side of it would be,
> >> unlike in the case of clocks where the devices assigning a default clock rate
> >> might be doing so on a device specific clock (rarely shared with other devices)
> >> in case of power domain, and especially in the qcom implementation of these
> >> power domains which support perf state, these can be large domains with lots of devices,
> >> and any device being active (not necessarily wanting any default perf state) will keep
> >> the domain at the default perf state, requested by a device which isn't really active.
> >
> > Yep, this certainly sounds suboptimal. To me, this isn't good enough.
> >
> >>
> >>> What about doing this just for the assigned-performance-state case as
> >>> the clients don't want to play with it at all.
> >>
> >> well, thats possible too, but you obviously can't reuse the same bindings
> >> in such cases
> >
> > Not sure I understand the issue with the DT binding? Let me elaborate
> > on how I think we could move forward.
> >
> > It looks like we have two problems to solve:
> >
> > *) We need a new DT binding.
> > If that becomes a generic property along the lines of the
> > "assigned-performance-state" as suggested - or if we decide to add a
> > SoC specific binding via using an additional cell in "power-domains"
> > (suggested by Rob), doesn't really matter much to me. The arguments
> > for the new DT property are very much similar to why we added
> > "assigned-clock-rates" for clocks.
> >
> > **) We want to avoid boiler-plate code in drivers to manage
> > "assigned-performance-state" for their devices.
> > No matter what DT property we decide on (generic or SoC specific), we
> > should be able to manage this from the PM domain (genpd) layer. No
> > changes in the drivers should be needed.
> > If a generic binding is used, we could consider to let genpd
> > internally manage the whole thing (DT parsing and updating performance
> > state votes for assigned-performance-state only).
>
> Sure, so for starters does that mean I should re-spin my series which
> adds the generic 'assigned-performance-states' bindings and see if Rob
> is OK with that? I am guessing you are OK with the way that binding gets
> used within genpd core in that series, or would you want it to be handled
> differently?
A re-spin would definitely move this forward. If we can convince Rob
about the generic DT binding, I will certainly agree to change genpd
to help/manage the DT parsing.
Although, I am not sure yet what is the best. Let genpd to do the
parsing internally in genpd_add_device() or just provide a helper
function that can be called from an ->attach|detach_dev() callback.
In the end it boils done to decide whether we should use the
->start|stop() callbacks or let genpd internally manage the "assigned
performance state". Honestly, I would prefer to look at how the code
would need to be changed, before answering this.
>
> > If we go for an SoC specific binding, the genpd provider needs to be
> > updated. It can manage DT parsing from the ->attach|detach_dev()
> > callbacks and update performance votes from the ->start|stop()
> > callbacks.
> > We could also consider a hybrid of these two solutions.
> >>
> >> [1] https://lore.kernel.org/patchwork/patch/1284042/
Kind regards
Uffe
