Armada 3720 cpufreq bug: request for testing (EspressoBIN, uDPU, ...)

[Pali Rohár]

On Monday 12 October 2020 07:19:48 Andre Heider wrote:
> Hi Marek,
> 
> On 09/10/2020 12:57, Marek Behún wrote:
> > I have a request for testing some patches that fix cpufreq issues on
> > Armada 3720 (EspressoBIN, uDPU, Devel Board, maybe others?).
> > 
> > It concerns top 4 commits of branch a3720-cpufreq-issues of repository
> > https://git.kernel.org/pub/scm/linux/kernel/git/kabel/linux.git/
> > 
> > (log at https://git.kernel.org/pub/scm/linux/kernel/git/kabel/linux.git/log/?h=a3720-cpufreq-issues)
> > 
> > We've discovered a bug in Armada 3720 cpufreq driver that causes
> > base CPU frequency to drop to 800 MHz on systems configured to 1 GHz.
> > More info here:
> > https://www.spinics.net/lists/arm-kernel/msg844752.html
> > 
> > The patchset above led us to discover another bug, in the SOC itself,
> > that causes the CPU to behave weirdly when frequency changes to 1 GHz.
> > A similar erratum is documented by Marvell but only for systems where
> > base frequency is 1.2 GHz.
> > 
> > We've discovered that to make cpufreq scaling stable on Armada 3720
> > on systems with base frequency 1 GHz, we also have to set voltage levels
> > for L1, L2 and L3 loads to 1108 mV. (We were led to this by patch we found here
> > https://forum.armbian.com/topic/10429-how-to-make-espressobin-v7-stable/page/2/?tab=comments#comment-80012).
> > 
> > We would like you to build this sources and test the ondemand cpufreq
> > governor on systems with variable CPU load, for at least 1 day (more is better).
> > If you could also test the userspace governor and randomly change frequency,
> > it would be great.
> > 
> > Please let me know if and when you are willing to do this.
> 
> Thanks for hunting this one down!
> 
> I've applied your 4 patches to v5.9 and gave it a quick spin.

Hello! I have there a new version of Armada 3720 CPU patches. They are
in branch "a3720-cpufreq-issues" at my git repository:

https://git.kernel.org/pub/scm/linux/kernel/git/pali/linux.git

Full commit log is here:
https://git.kernel.org/pub/scm/linux/kernel/git/pali/linux.git/log/?h=a3720-cpufreq-issues

The main difference from previous version of patches is that voltage
level for L2 and L3 loads is not increased, so power consumption for 200
and 250 MHz frequency should not be changed. Only for L1 load (500 MHz)
is increased to 1108 mV.

Could you test ondemand governor with these new patches?

> I'm using schedutil, which had the very same issue. It never actually
> switched to 1000mhz. It claimed it did, but the `mhz` util as well as the 7z
> benchmark proved it never used anything >800mhz.
> 
> Before:
> $ 7z b
> 
> 7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
> 
> p7zip Version 16.02 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,64 bits,2 CPUs
> LE)
> 
> LE
> CPU Freq:   793   793   794   793   794   792   793   794
> 
> 
> RAM size:     985 MB,  # CPU hardware threads:   2
> 
> RAM usage:    441 MB,  # Benchmark threads:      2
> 
> 
>                        Compressing  |                  Decompressing
> 
> Dict     Speed Usage    R/U Rating  |      Speed Usage    R/U Rating
> 
>          KiB/s     %   MIPS   MIPS  |      KiB/s     %   MIPS   MIPS
> 
> 
> 22:        711   149    464    692  |      17501   194    770   1494
> 
> 23:        708   155    465    722  |      17595   198    770   1523
> 
> 24:        682   158    464    734  |      17079   196    764   1499
> 
> 25:        663   160    473    757  |      15595   189    734   1388
> ----------------------------------  | ------------------------------
> Avr:             156    466    726  |              194    760   1476
> Tot:             175    613   1101
> 
> 
> After:
> $ 7z b
> 
> 7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
> p7zip Version 16.02 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,64 bits,2 CPUs
> LE)
> 
> LE
> CPU Freq:   992   998   998   994   996   997   997   997
> 
> RAM size:     987 MB,  # CPU hardware threads:   2
> RAM usage:    441 MB,  # Benchmark threads:      2
> 
>                        Compressing  |                  Decompressing
> Dict     Speed Usage    R/U Rating  |      Speed Usage    R/U Rating
>          KiB/s     %   MIPS   MIPS  |      KiB/s     %   MIPS   MIPS
> 
> 22:        866   151    559    843  |      22334   198    962   1907
> 23:        872   152    584    889  |      21809   197    959   1888
> 24:        808   145    599    869  |      11413   110    915   1002
> 25:        716   137    596    819  |      21092   198    949   1877
> ----------------------------------  | ------------------------------
> Avr:             146    585    855  |              176    946   1669
> Tot:             161    765   1262
> 
> 
> The measured frequency as well as the performance data finally looks good,
> so:
> 
> Tested-by: Andre Heider <a.heider at gmail.com>
> 
> Thanks,
> Andre