Re: [PATCH v7 0/4] qcom-cpufreq-hw: Add CPU clock provider support

From: Sudeep Holla <sudeep.holla@arm.com>
To: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>
Cc: andersson@kernel.org, viresh.kumar@linaro.org,
	Sudeep Holla <sudeep.holla@arm.com>,
	krzysztof.kozlowski+dt@linaro.org, rafael@kernel.org,
	robh+dt@kernel.org, johan@kernel.org, devicetree@vger.kernel.org,
	linux-arm-msm@vger.kernel.org, linux-kernel@vger.kernel.org,
	linux-pm@vger.kernel.org
Subject: Re: [PATCH v7 0/4] qcom-cpufreq-hw: Add CPU clock provider support
Date: Thu, 17 Nov 2022 13:23:19 +0000	[thread overview]
Message-ID: <20221117132319.cx7cwt3dwywlt6wx@bogus> (raw)
In-Reply-To: <20221117123841.GI93179@thinkpad>

On Thu, Nov 17, 2022 at 06:08:41PM +0530, Manivannan Sadhasivam wrote:

[...]
>
> AFAIK, OPP framework needs to know the current frequency of the device it is
> dealing with for setting the device's OPP. So it uses clk_get_rate() of the
> first clock of the device. If the clock is not available, then it uses the
> frequency in the first entry of the OPP table (since it is going to be the
> minimum freq of the device).
>

It has been a while since I followed OPP. Thanks for all the info, helped me
get updated without looking at the code in detail.

> As you can see, the clk_get_rate() is eminent for switching the OPPs and since
> OPP framework doesn't know what device it is dealing with, it cannot use
> cpufreq_get().
>

Agreed. I had assumed the qcom-cpufreq-hw as setting cpufreq directly pocking
the hardware but now I see it is using opp library to set some additional
policy.

> Is SCMI node itself has the OPP tables? Or the consumer nodes of the SCMI?
>

No, OPPs are read from the f/w and we just use OPP APIs to register them.
But we don't use OPP library to set the performance.

> TLDR; If you tell OPP framework to set a new OPP for a device, it needs to the
> know the current frequency of the device. And it is not manadatory now, but in
> the future maybe.
>

Hmm, good to know. I prefer it is not coupled with clocks and have some
alternative mechanism that is suitable for performance domains and don't
enforce the use of clock bindings and clk framework unnecessarily.

--
Regards,
Sudeep`