Re: [PATCH] intel_idle: Add Jasper Lake and Elkhart Lake support

[Rafael J. Wysocki]

On Thu, Aug 1, 2024 at 8:55 AM Kai-Heng Feng
<kai.heng.feng@xxxxxxxxxxxxx> wrote:
>
> On Thu, Aug 1, 2024 at 12:28 AM Len Brown <lenb@xxxxxxxxxx> wrote:
> >
> > On Wed, Jul 31, 2024 at 2:18 AM Kai-Heng Feng
> > <kai.heng.feng@xxxxxxxxxxxxx> wrote:
> > >
> > > The crucial part to make the issue (i.e. slow ethernet) is
> > > ".disable_promotion_to_c1e = true".
> >
> > Okay, so the problem statement is that on this machine with some
> > ethernet controller and some workload,
> > performance is better when you use just C1 and not C1E (or deeper) states.
> >
> > And so you want to have the option of accessing C1 without the overhead of C1E?
>
> Yes, that's the case here.
>
> >
> > Presumably you don't care about the power savings of the deeper states,
> > or you are using PM_QOS to avoid deep c-states at run time?
>
> I tried to use cpu_latency_qos in the network driver's NAPI poll, but
> only saw marginal improvement to around 830Mbps. Hitting the 940Mbps
> is still the goal here.
>
> >
> > > Can we use that for EHL and JSL?
> >
> > Yes.
>
> Is it plausible to disable C1E promotion while using ACPI idle driver?

It is in general.

It would require the driver to recognize the processors in question
without providing C-state tables, but it should be doable.

Next week I'll be offline, I'll look into this when I'm back.

> Or provide a C-state table in intel_idle and update the states via _CST?

That should be possible either.

> >
> > You may also have a BIOS option to achieve the same goal, depending on
> > the platform.)
>
> I am seeing three different platforms from different vendors hitting
> the same issue, so it's better to disable C1e for these platforms.

I see.

Thanks!