Fix spam error message when setting energy_performance_preference

[FosRexx]

The 'intel_pstate' driver does not allow the EPP to be set to anything but 'performance' when the scaling governor is set to 'performance', previously auto-cpufreq when the scaling-governor was set to 'performance' tried to set the EPP to 'balance-performance' which caused a spam of write error messages in journalctl in system with 'intel_pstate' drivers. This is an intended behavior, since according to the kernel documentation when HWP is enabled(which is enabled by default during boot with supported processors) and scaling governor is set to performance the processor’s internal P-state selection logic is expected to focus entirely on performance. And this will override the EPP setting and reject any value different from 0 (“performance”). This commit just changes the 'balance-performance' EPP preference in set_performance() to 'performance'. Which fixes the spam issue.

Fixes: #661

[FosRexx]

In the original issue #661 someone commented having the same issue with an AMD CPU with amd-pstate-epp driver and again in this closed issue #563 the same exact issue is reported for AMD CPU with amd-pstate-epp driver, which leads me to believe that this behavior happens in both AMD and Intel CPU with EPP support. I cannot confirm that this will work for the amd-pstate-epp drivers and I could not find anything in the amd-pstate kernel documentation on this matter. So, I am hesitated to only apply this fix for intel-pstate drivers, since this could also fix the issue for amd-pstate-epp drivers, and the availability of EPP is already checked in the code before setting the energy_performance_preference so we don't have to worry about this applying for non-epp drivers. It would be great if someone using the amd-pstate-epp driver could verify this fix.

And yes this does not occur with set_powersave(), since in this mode the scaling governor is set to powersave and in this case their is no limitation to energy_performance_perference as stated here. This behavior of not allowing the energy_performance_preference to be set to anything but performance at-least for the intel-pstate driver is intended and mentioned in the kernel documentation. The HWP feature is enabled by default for processors supporting it as mentioned here.

[AdnanHodzic]

I absolutely agree with what was said here and share the same opinion:

Thanks for finding the cause of the write error spam. but I think we want to keep balance_performance as the default EPP, so the solution to this would be to check if the user is running intel_pstate, if they are, set the EPP to performance, otherwise set it to balance_performance

Also, since I'm using intel_pstate driver am I misunderstanding something because when I'm on powersave governor, EPP will be set to power EPP?

The 'intel_pstate' driver does not allow the EPP to be set to anything but 'performance'

i.e:

Setting to use: "powersave" governor
Setting to use: "power" EPP

and opposite will be the case when setting performance governor, i.e:

Setting to use: "performance" governor
Setting to use: "performance" EPP

It would be great if someone using the amd-pstate-epp driver could verify this fix.

Aren't you running on AMD @shadeyg56 ?

[FosRexx]

First, let me resolve the misunderstanding which was caused due to missing information in the PR comment and the commit message, which I have now hopefully resolved and new people stumbling on this PR will understand this better.

Also, since I'm using intel_pstate driver am I misunderstanding something because when I'm on powersave governor, EPP will be set to power EPP?

So, this is the relevant section of the kernel documentation regarding the why's of the issue. In summary, in system's with Intel CPU and AMD CPU, that support the hardware-managed P-states (HWP) running the intel_pstate driver in active mode(which is the default), when the scaling governor is set to performance, the driver does not allow the energy-performance-preference to be set to any value different than performance. If the value is set to any value different than performance via sysfs interface, that configuration will be rejected.

This behavior causes the error message "write error: Device or resource busy", when auto-cpufreq tries to set the EPP to balanced-performance when the scaling governor is set to performance i.e. in set_performance() function.

This does not happen when the scaling governor is set to powersave, since in this case the driver allows the energy-performance-preference value to to be set to any value. That's why when you are on powersave governor, the EPP is set to power(which is weird, might I add, since in the latest commit 8bb7478 in master the EPP should be set to balance_power when in powersave governor, see function set_powersave())

Hopefully this clears the misunderstanding.

Regarding keeping the default balance_performance as the default EPP.

but I think we want to keep balance_performance as the default EPP, so the solution to this would be to check if the user is running intel_pstate, if they are, set the EPP to performance, otherwise set it to balance_performance

I am hesitated to only apply this fix for intel_pstate, because according to this comment and currently closed issue #563, this issue also occurs in amd_pstate_epp driver. I do not have the first hand experience of this issue in AMD and I could not find anything regarding this on the amd_pstate kernel documentation, but as you can see from the comment and the Issue this also happens in AMD systems. And so, this PR might also fix the issue for amd_pstate_epp users, and thats why I am hesitant(I do not know if this expression makes any sense, as you might have already guessed English is not my first language). And so all the system with EPP are affected and non-EPP users should not be affected by this change in code, since the check for if the system has EPP or not is already been done.

[FosRexx]

If you insist on only applying this fix for intel_pstate drivers and keeping the default EPP to balanced_performance then I have update the code to your liking in 828db0ef31d13bc299f143472b2e60a57129eb5b commit.

[shadeyg56]

I am hesitated to only apply this fix for intel_pstate, because according to this comment and currently closed issue #563, this issue also occurs in amd_pstate_epp driver. I do not have the first hand experience of this issue in AMD and I could not find anything regarding this on the amd_pstate kernel documentation

I've used amd_pstate_epp before and can confirm that it supports balance_performance. Also based on these benchmarks, this shows amd can use balance_performance. I believe the issue you linked had a different cause for AMD systems. Currently, I'm not getting any error spam on my AMD system

[FosRexx]

Well, I look like an idiot now 🥲, I already updated the code to only apply this for intel_pstate driver.

[parmjotsinghrobot]

Also does this not occur with set_powersave() as well since that uses balance_power? I can't test any of this as I don't have an Intel CPU

Yes it does happen with Intel, I have tested it with the powersave governor and the power EPP has to be used with it for the error to not be spammed in the logs.

I think before merging, the MR should be changed to also include this case, but I do believe the way that this is done is not the best. I think that instead of just checking if the path exists, we should read the value of said path and set the value accordingly.

[AdnanHodzic]

I think before merging, the MR should be changed to also include this case, but I do believe the way that this is done is not the best. I think that instead of just checking if the path exists, we should read the value of said path and set the value accordingly.

+1

[FosRexx]

Yes it does happen with Intel, I have tested it with the powersave governor and the power EPP has to be used with it for the error to not be spammed in the logs.

Cannot reproduce

OS: Arch Linux x86_64 Host: Nitro AN515-58 (V2.10) Kernel: 6.8.7-arch1-2 CPU: 12th Gen Intel(R) Core(TM) i5-12450H (12) @ 4.40 GHz

[parmjotsinghrobot]

Yes it does happen with Intel, I have tested it with the powersave governor and the power EPP has to be used with it for the error to not be spammed in the logs.

Cannot reproduce

Strange. After a kernel update I cannot reproduce either. This means that this might be an upstream kernel bug but what's the point of waiting for them.

I like your current change regarding checking the state. @AdnanHodzic please review.

[FosRexx]

Updated the code so that user written config energy_performance_preference are also handled.

[AdnanHodzic]

Updated the code so that user written config energy_performance_preference are also handled.

I started testing the changes on this branch, removed my auto-cpufreq config file and wanted to see what it runs on "default" settings. While laptop is connected to power it all works as expected in terms of EPP and governors.

But then, when pulled out the power cable and ran: stress --cpu 8 --io 4 --vm 4 --timeout 15

I noticed that both governor and EPP remained pinned to "powersave", and what's more worrying my CPU cores were also pinned to max 800Mhz. As while auto-cpufreq will switch you to powersave when when connected to the charger, it should switch over to performance if it detects such high cpu usage/load.

Linux distro: Ubuntu 24.04 Noble Numbat
Linux kernel: 6.8.0-31-generic
Processor: Intel(R) Core(TM) i7-8565U CPU @ 1.80GHz
Cores: 8
Architecture: x86_64
Driver: intel_pstate

------------------------------ Current CPU stats ------------------------------

CPU max frequency: 800 MHz
CPU min frequency: 400 MHz

Core    Usage   Temperature Frequency
CPU0      2.1%        48 °C       800 MHz
CPU1     12.1%        48 °C       800 MHz
CPU2      8.2%        47 °C       765 MHz
CPU3     52.0%        49 °C       400 MHz
CPU4      3.0%        48 °C       400 MHz
CPU5      4.0%        48 °C       800 MHz
CPU6      0.0%        47 °C       800 MHz
CPU7     15.3%        49 °C       400 MHz

CPU fan speed: 0 RPM

---------------------------- CPU frequency scaling ----------------------------

Battery is: discharging

Setting to use: "powersave" governor
Setting to use: "power" EPP

Total CPU usage: 100 %
Total system load: 3.99
Average temp. of all cores: 48.00 °C 

Load optimal (load average: 3.99, 1.90, 1.37)
setting turbo boost: off

But then I noticed this was the case on master branch as well. Can anyone else reproduce this (i.e @shadeyg56 & @PurpleWazard), as if this is truly the case we had some kind of regression along the way.

[AdnanHodzic]

Please disregard my last reply, from code POV, changes look good now, and I've been running auto-cpufreq with them for couple of days and I had 0 problems.

@shadeyg56 I requested re-review of comment I made, as from what I see this request has been fulfilled, if not please raise it.

Otherwise, thank you for your contribution @FosRexx, you will be credited for your work as part of upcoming auto-cpufreq release!