Message-ID: <83791a71-a45c-383d-0406-b0f4e0a0c215@baylibre.com>
Date:   Fri, 13 Dec 2019 14:24:24 +0100
From:   Neil Armstrong <narmstrong@...libre.com>
To:     Anand Moon <linux.amoon@...il.com>,
        Martin Blumenstingl <martin.blumenstingl@...glemail.com>
Cc:     Kevin Hilman <khilman@...libre.com>,
        Rob Herring <robh+dt@...nel.org>,
        Jerome Brunet <jbrunet@...libre.com>,
        devicetree <devicetree@...r.kernel.org>,
        linux-arm-kernel <linux-arm-kernel@...ts.infradead.org>,
        linux-amlogic@...ts.infradead.org,
        Linux Kernel <linux-kernel@...r.kernel.org>
Subject: Re: [RFC-next 0/1] Odroid C2: Enable DVFS for cpu

On 13/12/2019 12:28, Anand Moon wrote:
> Hi Martin
> 
> On Fri, 13 Dec 2019 at 01:40, Martin Blumenstingl
> <martin.blumenstingl@...glemail.com> wrote:
>>
>> Hi Neil,
>>
>> On Wed, Dec 11, 2019 at 9:49 AM Neil Armstrong <narmstrong@...libre.com> wrote:
>>>
>>> On 10/12/2019 22:47, Kevin Hilman wrote:
>>>> Martin Blumenstingl <martin.blumenstingl@...glemail.com> writes:
>>>>
>>>>> On Tue, Dec 10, 2019 at 7:13 PM Kevin Hilman <khilman@...libre.com> wrote:
>>>>>>
>>>>>> Anand Moon <linux.amoon@...il.com> writes:
>>>>>>
>>>>>>> Hi Neil / Kevin,
>>>>>>>
>>>>>>> On Tue, 10 Dec 2019 at 14:13, Neil Armstrong <narmstrong@...libre.com> wrote:
>>>>>>>>
>>>>>>>> On 09/12/2019 23:12, Kevin Hilman wrote:
>>>>>>>>> Anand Moon <linux.amoon@...il.com> writes:
>>>>>>>>>
>>>>>>>>>> Some how this patch got lost, so resend this again.
>>>>>>>>>>
>>>>>>>>>> [0] https://patchwork.kernel.org/patch/11136545/
>>>>>>>>>>
>>>>>>>>>> This patch enable DVFS on GXBB Odroid C2.
>>>>>>>>>>
>>>>>>>>>> DVFS has been tested by running the arm64 cpuburn
>>>>>>>>>> [1] https://github.com/ssvb/cpuburn-arm/blob/master/cpuburn-a53.S
>>>>>>>>>> PM-QA testing
>>>>>>>>>> [2] https://git.linaro.org/power/pm-qa.git [cpufreq testcase]
>>>>>>>>>>
>>>>>>>>>> Tested on latest U-Boot 2019.07-1 (Aug 01 2019 - 23:58:01 +0000) Arch Linux ARM
>>>>>>>>>
>>>>>>>>> Have you tested with the Harkernel u-boot?
>>>>>>>>>
>>>>>>>>> Last I remember, enabling CPUfreq will cause system hangs with the
>>>>>>>>> Hardkernel u-boot because of improperly enabled frequencies, so I'm not
>>>>>>>>> terribly inclined to merge this patch.
>>>>>>>
>>>>>>> HK u-boot have many issue with loading the kernel, with load address
>>>>>>> *it's really hard to build the kernel for HK u-boot*,
>>>>>>> to get the configuration correctly.
>>>>>>>
>>>>>>> Well I have tested with mainline u-boot with latest ATF .
>>>>>>> I would prefer mainline u-boot for all the Amlogic SBC, since
>>>>>>> they sync with latest driver changes.
>>>>>>
>>>>>> Yes, we would all prefer mainline u-boot, but the mainline kernel needs
>>>>>> to support the vendor u-boot that is shipping with the boards.  So
>>>>>> until Hardkernel (and other vendors) switch to mainline u-boot we do not
>>>>>> want to have upstream kernel defaults that will not boot with the vendor
>>>>>> u-boot.
>>>>>>
>>>>>> We can always support these features, but they just cannot be enabled
>>>>>> by default.
>>>>> (I don't have an Odroid-C2 but I'm curious)
>>>>> should Anand submit a patch to mainline u-boot instead?
>>>>
>>>> It would be in addition to $SUBJECT patch, not instead, I think.
>>>>
>>>>> the &scpi_clocks node could be enabled at runtime by mainline u-boot
>>>>
>>>> That would work, but I don't know about u-boot maintainers opinions on
>>>> this kind of thing, so let's see what Neil thinks.
>>>
>>> U-Boot doesn't anything to do with SCPI, SCPI discusses directly with the SCP
>>> processor, and the CPU clock is set to 1,56GHz by the BL2 boot stage before
>>> U-boot starts.
>>>
>>> The only viable solution I see now is to find if we could add a DT OPP table
>>> only for Odroid-C2 dts to bypass the SCPI OPP table.
>> my understanding is that mainline u-boot (with whatever SCP firmware
>> it uses) provides the *correct* OPP table
> 
> Right now I am not sure how this OPP table is populated.
> But I saw the same freq table used in 3.16.x kernel after enable the clk.
> 
>> in this case it would be "safe" to have SCPI enabled with mainline u-boot
>> @Anand: please correct me if I misunderstood you
>>
> 
> As per my understanding DVFS OPP frequency table for SCPI firmware set
> for 1.536 GHz
> somewhere in BL2 as pointed by Neil.
> 
> Arm Trusted firmware added new secure SCPI communication with
> Cortex-M3 co processor.
> [0] https://github.com/ARM-software/arm-trusted-firmware/blob/master/docs/plat/meson-gxbb.rst
> [1] https://github.com/ARM-software/arm-trusted-firmware/blob/master/plat/amlogic/common/aml_scpi.c
> 
> ATF generated the *bl1.bin* which is replace the Amlogic's bl1.bin
> while preparing
> the new u-boot *u-boot.gxbb* image.
> 
>> my idea to "enable SCPI with mainline u-boot" is to have u-boot update
>> the "status" property of the scpi_clocks node.
>> u-boot does something similar with the mac-address property of the
>> Ethernet controller for example.
>> as result of this users of mainline u-boot would have working CPU
>> DVFS, while users of the old vendor u-boot would run at fixed 1.54GHz.
>>
>>
>> Martin
> 
> Right now as per my understanding 1.536 GHz max is bit under clocked.
> 
> Some time ago on Odroid Forum tried to over clock the cpu to 2GHz.

This is the point, the Odroid-C2 is *not* stable at 2GHz,
a large amount of board doesn't support 2GHz, this is why Amlogic
dropped the freq > 1.536 GHz for the GXBB family.

But HardKernel still delivers the SCPI table with > 1.536 GHz which breaks
on most of the boards, but doesn't on 3.14 since they have a hack disabling
higher freqs with a cmdline set in boot.ini.

> [3] https://forum.odroid.com/viewtopic.php?f=139&t=18738
> So more investigation need to done in this line.
> 
> I also tried the same with HardKernel Image, with modifying the boot.ini
> I could increase the max DVFS cpu frequency to 1.90 GHz,
> This is just proof of concept.
> 
> odroid:~# cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_available_frequencies
> 100000 250000 500000 1000000 1296000 1536000 1656000 1680000 1752000 1896000
> 
> I have some minimal stress testing attached are the results for HK
> 3.16.x kernel.
> 
> For now we should not enable this clock.
> Until we can possible to check for higher clock frequency to work stable
> on all Amlogic S905X SBC.
> 
> I like the Neil's approach to use it's own dts OPP table for SCPI protocol.

The various tests gave very little perf enhancement by going > 1,536 GHz, seriously
it's not worth the pain.

Neil

> 
> -Anand
> 

Powered by blists - more mailing lists