Some AMD GPUs Affected By A Nasty Power Regression That Snuck Into Linux 4.18 Stable

[nuetzel]

Interesting. I have to wonder if this impacts laptop chipset implementation. I have this feeling that the latest Fedora Kernels have shortened my battery life. The problem is no objective testing so maybe it is me.

Show us `cat /sys/kernel/debug/dri/0/amdgpu_pm_info` like Alex Deucher suggested in the kernel bug tracker. Maybe even before and after your findings.

BTW I'm off for family vacation after Saturday morning.
Good night!

[tuxd3v]

Do you like doing 'double work'...;-)

Little search (or advice to 'reporter' to search) in https://bugzilla.kernel.org or https://bugs.freedesktop.org had yield that it was already reported...

Greetings,
Dieter

That is very unfair..to say at least..

If it was not Michael, we out there would be in the dark with this problem!!
Because of that, we know whats going on, and the some root events that lead to the problem.

He made his Job very well, _now_, Amd needs to do their Job..

[nuetzel]

That is very unfair..to say at least..

If it was not Michael, we out there would be in the dark with this problem!!
Because of that, we know whats going on, and the some root events that lead to the problem.

He made his Job very well, _now_, Amd needs to do their Job..

I think he know how it was meant!
All _his_ bisect was done so he had 'only' write about it and saved some work and time...

[tomtomme]

Maybe AMD / Intel / Distros should default install LTS kernels for their GPUs except maybe for the last/newest gen?

I had so many problems with 4.17/18 kernels, that after many years I switched from openSUSE Tumbleweed to Manjaro because it comes with an LTS kernel per default (I know about kernel:HEAD repo on suse).
And boom, all problems were gone. Multimonitor without hicups, suspend / resume worked again, 4k@60hz through my hdmi/dp adapter worked again, booting without safe mode on the same adapter worked. This is on an Radeon Fury Desktop and an Haswell Iris Laptop.
With Tumbleweed releasing 1 kernel per week, there was just too much of an unpredictable change of bugs, while a whole LTS Distro on the other side would be too slow, on the desktop software side for my taste.

[dwagner]

Hello dwagner,

do NOT help:

SunWave1 card0/device# echo 0 > pp_dpm_mclk
SunWave1 card0/device# echo 0 > pp_dpm_sclk
SunWave1 card0/device# cat pp_dpm_mclk
0: 300Mhz
1: 1000Mhz
2: 2000Mhz *
SunWave1 card0/device# cat pp_dpm_sclk
0: 300Mhz
1: 600Mhz
2: 900Mhz
3: 1145Mhz
4: 1215Mhz
5: 1257Mhz
6: 1300Mhz
7: 1411Mhz *

Did you not mention or not enter the "echo manual >power_dpm_force_performance_level"? Without it, your writes to pp_dpm_* are ignored.

[stingray454]

Hm, a few weeks ago my monitor started glitching / "crackling" after turning it off and on again (by power settings or manually), but only in 60hz - turned it down to 50 and all problems were gone. Timing wise it seems to be around the time for this "fix" I started noticing it, I should really look into if it's related.

[smirky]

Wow, my monitor also started turning off and on since recent kernel versions. I though it was DM/WM related or some power settings not working correctly. Polaris RX480 + Philips 234CL monitor.

[ptyerman]

Currently I'm in Windows and GPU-Z is reporting a stable 28.4W at idle. This is a Sapphire RX580 8Gb model card.

[Strunkenbold]

Yeah I don't get it either, for such recent and common cards it should be caught during QA.

I think their QA is automated tests. If there is no test, to test idle power consumption, it doesnt get caught.
Further I think there is no one at AMD doing extensive testing apart from these automated tests. They are just too busy.
Given that AMD has to implement many features in the Linux Kernel to reach to the same feature level like the Windows drivers, I guess minimal testing and thus resulting in many regressions, are expected.
I really would hope AMD does a little more for driver quality. There are a plethora of crash reports on the bug tracker but usually no reaction. Maybe they should think about contracting a consulting company to address those bugs one by one.

[Brisse]

Sapphire Nitro R9 Fury on Debian Sid. No issue. Reading the comments has me thinking this is exclusive to Polaris cards.

uname -a; sensors | grep -A 5 amdgpu
Linux kristoffer-debian-desktop 4.18.0-1-amd64 #1 SMP Debian 4.18.8-1 (2018-09-18) x86_64 GNU/Linux
amdgpu-pci-0900
Adapter: PCI adapter
vddgfx:       +0.85 V  
fan1:         766 RPM
temp1:        +54.0°C  (crit = +89.0°C, hyst = -273.1°C)
power1:       14.17 W  (cap = 100.00 W)

Note: Fans are not actually spinning. This is just their last reported RPM before shutting down. Also, power cap is normally 260W, not 100W. I changed it because I was mining just before.