Re: [PATCH v2 4/4] cpufreq: mediatek: Raise proc and sram max voltage for MT7622/7623

[AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>]

Il 22/05/23 20:03, Daniel Golle ha scritto:
> Hi Jia-Wei,
> Hi AngeloGioacchino,
> 
> On Fri, Mar 24, 2023 at 06:11:30PM +0800, jia-wei.chang wrote:
>> From: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
>>
>> During the addition of SRAM voltage tracking for CCI scaling, this
>> driver got some voltage limits set for the vtrack algorithm: these
>> were moved to platform data first, then enforced in a later commit
>> 6a17b3876bc8 ("cpufreq: mediatek: Refine mtk_cpufreq_voltage_tracking()")
>> using these as max values for the regulator_set_voltage() calls.
>>
>> In this case, the vsram/vproc constraints for MT7622 and MT7623
>> were supposed to be the same as MT2701 (and a number of other SoCs),
>> but that turned out to be a mistake because the aforementioned two
>> SoCs' maximum voltage for both VPROC and VPROC_SRAM is 1.36V.
>>
>> Fix that by adding new platform data for MT7622/7623 declaring the
>> right {proc,sram}_max_volt parameter.
>>
>> Fixes: ead858bd128d ("cpufreq: mediatek: Move voltage limits to platform data")
>> Fixes: 6a17b3876bc8 ("cpufreq: mediatek: Refine mtk_cpufreq_voltage_tracking()")
>> Signed-off-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
>> Signed-off-by: Jia-Wei Chang <jia-wei.chang@mediatek.com>
>> ---
>>   drivers/cpufreq/mediatek-cpufreq.c | 13 +++++++++++--
>>   1 file changed, 11 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/cpufreq/mediatek-cpufreq.c b/drivers/cpufreq/mediatek-cpufreq.c
>> index 764e4fbdd536..9a39a7ccfae9 100644
>> --- a/drivers/cpufreq/mediatek-cpufreq.c
>> +++ b/drivers/cpufreq/mediatek-cpufreq.c
>> @@ -693,6 +693,15 @@ static const struct mtk_cpufreq_platform_data mt2701_platform_data = {
>>   	.ccifreq_supported = false,
>>   };
>>   
>> +static const struct mtk_cpufreq_platform_data mt7622_platform_data = {
>> +	.min_volt_shift = 100000,
>> +	.max_volt_shift = 200000,
>> +	.proc_max_volt = 1360000,
>> +	.sram_min_volt = 0,
>> +	.sram_max_volt = 1360000,
> 
> This change breaks cpufreq (with ondemand scheduler) on my BPi R64
> board (having MT7622AV SoC with MT6380N PMIC).
> ...
> [    2.540091] cpufreq: __target_index: Failed to change cpu frequency: -22
> [    2.556985] cpu cpu0: cpu0: failed to scale up voltage!
> ...
> (repeating a lot, every time the highest operating point is selected
> by the cpufreq governor)
> 
> The reason is that the MT6380N doesn't support 1360000uV on the supply
> outputs used for SRAM and processor.
> 
> As for some reason cpufreq-mediatek tries to rise the SRAM supply
> voltage to the maximum for a short moment (probably a side-effect of
> the voltage tracking algorithm), this fails because the PMIC only
> supports up to 1350000uV. As the highest operating point is anyway
> using only 1310000uV the simple fix is setting 1350000uV as the maximum
> instead for both proc_max_volt and sram_max_volt.
> 
> A similar situation applies also for BPi R2 (MT7623NI with MT6323L
> PMIC), here the maximum supported voltage of the PMIC which also only
> supports up to 1350000uV, and the SoC having its highest operating
> voltage defined at 1300000uV.
> 
> If all agree with the simple fix I will post a patch for that.
> 
> However, to me it feels fishy to begin with that the tracking algorithm
> tries to rise the voltage above the highest operating point defined in
> device tree, see here:
> 
> 6a17b3876bc830 drivers/cpufreq/mediatek-cpufreq.c (Jia-Wei Chang              2022-05-05 19:52:20 +0800 100)    new_vsram = clamp(new_vproc + soc_data->min_volt_shift,
> 6a17b3876bc830 drivers/cpufreq/mediatek-cpufreq.c (Jia-Wei Chang              2022-05-05 19:52:20 +0800 101)                      soc_data->sram_min_volt, soc_data->sram_max_volt);
> 
> However, I did not investigate in depth the purpose of this
> initial rise and can impossibly test my modifications to the
> tracking algorithm on all supported SoCs.
> 

Thanks for actually reporting that, I don't think that there's any
valid reason why the algorithm should set a voltage higher than the
maximum votage specified in the fastest OPP.

Anyway - the logic for the platform data of this driver is to declare
the maximum voltage that SoC model X supports, regardless of the actual
board-specific OPPs, so that part is right; to solve this issue, I guess
that the only way is for this driver to parse the OPPs during .probe()
and then always use in the algorithm

	vproc_max = max(proc_max_volt, opp_vproc_max);
	vsram_max = max(sram_max_volt, vsram_vreg_max);

Jia-Wei, can you please handle this?

Thanks,
Angelo