Re: [PATCH 04/19] cpufreq: amd: introduce a new amd pstate driver to support future processors

[Huang Rui]

On Thu, Sep 16, 2021 at 07:19:21PM +0800, Rafael J. Wysocki wrote:
> On Thu, Sep 16, 2021 at 12:09 PM Huang Rui <ray.huang@xxxxxxx> wrote:
> >
> > On Mon, Sep 13, 2021 at 07:56:32PM +0800, Peter Zijlstra wrote:
> > > On Mon, Sep 13, 2021 at 06:54:58PM +0800, Huang Rui wrote:
> > > > On Mon, Sep 13, 2021 at 04:56:24PM +0800, Peter Zijlstra wrote:
> > >
> > > > > > 1) Full MSR Support
> > > > > > If current hardware has the full MSR support, we register "pstate_funcs"
> > > > > > callback functions to implement the MSR operations to control the clocks.
> > > > >
> > > > > What's the WRMSR cost for those? I've not really kept track of the MSR
> > > > > costs on AMD platforms, but on Intel it has (luckily) been coming down
> > > > > quite a bit.
> > > >
> > > > Good to know this, I didn't have a chance to give a check. May I know how
> > > > did you test this latency? But MSR is new hardware design for this
> > > > solution, as designer mentioned, the WRMSR is low-latency register model is
> > > > faster than ACPI AML code interpreter.
> > > >
> > > > >
> > > > > > 2) Shared Memory Support
> > > > > > If current hardware doesn't have the full MSR support, that means it only
> > > > > > provides share memory support. We will leverage APIs in cppc_acpi libs with
> > > > > > "cppc_funcs" to implement the target function for the frequency control.
> > > > >
> > > > > Right, the mailbox thing. How is the performance of this vs MSR accesses?
> > > >
> > > > I will give a check. If you have a existing test method that can be used, I
> > > > can check it quickly.
> > >
> > > Oh, I was mostly wondering if using the mailbox as MMIO would be faster
> > > than an MSR, but you've already answered that above. Also:
> > >
> > > > > > 1. As mentioned above, amd-pstate driver can implement
> > > > > > fast_switch/adjust_perf function with full MSR operations that have better
> > > > > > performance for schedutil and other governors.
> > > > >
> > > > > Why couldn't the existing cppc-cpufreq grow this?
> > > >
> > > > Because fast_switch can adjust the frequency directly in the interrupt
> > > > context, if we use the acpi cppc handling with shared memory solution, it
> > > > will have a deadlock. So fast switch needs the control with registers
> > > > directly like acpi-cpufreq and intel-pstate.
> > >
> > > Aah, I see, you're only doing fast_switch support when you have MSRs.
> > > That was totally non-obvious.. :/
> >
> > Yes, I should have written a comment to there. :-)
> > Will update this in V2.
> >
> > >
> > > But then amd_pstate_adjust_perf() could just direct call the pstate
> > > methods and we don't need that indirection *at*all*, right?
> >
> > Hmm, yes, if we use amd_pstate_adjust_perf here, we won't need to call
> > amd_pstate_fast_switch. I saw intel_pstate had adjust_perf and fast_switch
> > at the same time, would you mind to let me know how to distinguish these
> > two use scenario on intel processors?
> 
> The ->fast_switch() callback is for the use cases in which
> ->adjust_perf() cannot be installed, that is basically systems without
> HWP enabled.

OK, I see. Thanks to clarify this.

Thanks,
Ray