Message-ID: <7968db3b-8545-0601-4302-301a4006f3bc@linaro.org>
Date:   Thu, 2 Feb 2023 23:35:37 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.7.0
Subject: Re: [PATCH] cpufreq: qcom-hw: Fix cpufreq_driver->get() for non-LMH
 systems
Content-Language: en-US
To:     Douglas Anderson <dianders@chromium.org>,
        Viresh Kumar <viresh.kumar@linaro.org>,
        Manivannan Sadhasivam <mani@kernel.org>
Cc:     swboyd@chromium.org, mka@chromium.org,
        Andy Gross <agross@kernel.org>,
        Bjorn Andersson <andersson@kernel.org>,
        "Rafael J. Wysocki" <rafael@kernel.org>,
        Sasha Levin <sashal@kernel.org>, linux-arm-msm@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-pm@vger.kernel.org
References: <20230202140005.1.I4b30aaa027c73372ec4068cc0f0dc665af8b938d@changeid>
From:   Konrad Dybcio <konrad.dybcio@linaro.org>
In-Reply-To: <20230202140005.1.I4b30aaa027c73372ec4068cc0f0dc665af8b938d@changeid>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 2.02.2023 23:00, Douglas Anderson wrote:
> On a sc7180-based Chromebook, when I go to
> /sys/devices/system/cpu/cpu0/cpufreq I can see:
> 
>   cpuinfo_cur_freq:2995200
>   cpuinfo_max_freq:1804800
>   scaling_available_frequencies:300000 576000 ... 1708800 1804800
>   scaling_cur_freq:1804800
>   scaling_max_freq:1804800
> 
> As you can see the `cpuinfo_cur_freq` is bogus. It turns out that this
> bogus info started showing up as of commit 205f5e984d30 ("cpufreq:
> qcom-hw: Fix the frequency returned by cpufreq_driver->get()"). That
> commit seems to assume that everyone is on the LMH bandwagon, but
> sc7180 isn't.
> 
> Let's go back to the old code in the case where LMH isn't used.
> 
> Fixes: 205f5e984d30 ("cpufreq: qcom-hw: Fix the frequency returned by cpufreq_driver->get()")
> Signed-off-by: Douglas Anderson <dianders@chromium.org>
> ---
Actually I hit the exact same issue when working on CPRh-aware
cpufreq with manual OSM programming.. LMh gets enabled by the firmware
on most recent platforms, but it's not the case for some old-timers.

I figured that adding a bool broken_lmh_freq in driver data would be
a good middleground between reverting that patch and ignoring the
issue, because it *does* matter what this function reports on LMh-
enabled platforms (yes, the subsystems are bluepilled between each
other and OSM/EPSS does not know the *real* throttled frequency),
but obviously we don't want to report 2.99Ghz otherwise..

I think 7280 had an issue where a SoC-specific compatible was not
introduced when the DT part was first merged, same goes for 6115.
6115 does have firmware-enabled LMh, not sure about 7280. In case
you wanted to go that route, I think it would be suitable to add
a blacklist of retroactively-broken platforms (match-by-machine-
compatible; don't scream at me bindings folks, I guess that's the
least messy solution) in addition to either matching the SoC-specific
compatible to epss_broken_lmh_driver_data.

Or we can forget about old DTs and just bind qcom,sc7180-cpufreq-hw
(and 7280, maybe? please check.) to this new driver data without
checking the machine compatible.


Konrad
> 
>  drivers/cpufreq/qcom-cpufreq-hw.c | 24 +++++++++++++-----------
>  1 file changed, 13 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/cpufreq/qcom-cpufreq-hw.c b/drivers/cpufreq/qcom-cpufreq-hw.c
> index 9505a812d6a1..957cf6bb8c05 100644
> --- a/drivers/cpufreq/qcom-cpufreq-hw.c
> +++ b/drivers/cpufreq/qcom-cpufreq-hw.c
> @@ -143,40 +143,42 @@ static unsigned long qcom_lmh_get_throttle_freq(struct qcom_cpufreq_data *data)
>  	return lval * xo_rate;
>  }
>  
> -/* Get the current frequency of the CPU (after throttling) */
> -static unsigned int qcom_cpufreq_hw_get(unsigned int cpu)
> +/* Get the frequency requested by the cpufreq core for the CPU */
> +static unsigned int qcom_cpufreq_get_freq(unsigned int cpu)
>  {
>  	struct qcom_cpufreq_data *data;
> +	const struct qcom_cpufreq_soc_data *soc_data;
>  	struct cpufreq_policy *policy;
> +	unsigned int index;
>  
>  	policy = cpufreq_cpu_get_raw(cpu);
>  	if (!policy)
>  		return 0;
>  
>  	data = policy->driver_data;
> +	soc_data = qcom_cpufreq.soc_data;
>  
> -	return qcom_lmh_get_throttle_freq(data) / HZ_PER_KHZ;
> +	index = readl_relaxed(data->base + soc_data->reg_perf_state);
> +	index = min(index, LUT_MAX_ENTRIES - 1);
> +
> +	return policy->freq_table[index].frequency;
>  }
>  
> -/* Get the frequency requested by the cpufreq core for the CPU */
> -static unsigned int qcom_cpufreq_get_freq(unsigned int cpu)
> +static unsigned int qcom_cpufreq_hw_get(unsigned int cpu)
>  {
>  	struct qcom_cpufreq_data *data;
> -	const struct qcom_cpufreq_soc_data *soc_data;
>  	struct cpufreq_policy *policy;
> -	unsigned int index;
>  
>  	policy = cpufreq_cpu_get_raw(cpu);
>  	if (!policy)
>  		return 0;
>  
>  	data = policy->driver_data;
> -	soc_data = qcom_cpufreq.soc_data;
>  
> -	index = readl_relaxed(data->base + soc_data->reg_perf_state);
> -	index = min(index, LUT_MAX_ENTRIES - 1);
> +	if (data->throttle_irq >= 0)
> +		return qcom_lmh_get_throttle_freq(data) / HZ_PER_KHZ;
>  
> -	return policy->freq_table[index].frequency;
> +	return qcom_cpufreq_get_freq(cpu);
>  }
>  
>  static unsigned int qcom_cpufreq_hw_fast_switch(struct cpufreq_policy *policy,