AMDGPU graphics reset/crash when RADEON_DPM_PERF_LEVEL_ON_AC is set to high

[acarasimon96]

• I've read and accepted the Bug Reporting Howto
• I've attached all required tlp-stat outputs via Gist (see below)

Describe the bug

After upgrading TLP to 1.4.0, my ThinkPad laptop would stop responding a few seconds after I plug it in to AC power while RADEON_DPM_PERF_LEVEL_ON_AC is set to high in the config file. The display would freeze and sometimes show corruption artifacts. In extreme cases, the display would blank out, and my system would reboot without warning.

Steps to reproduce

1. Make sure that the installed version of TLP is 1.4.0, and that it is enabled in the tlp.conf and running.
2. Change the value of RADEON_DPM_PERF_LEVEL_ON_AC to "high" (with quotes) in tlp.conf and restart TLP
3. If you're on a TTY, switch to a desktop session (you don't have to log in)
4. Plug in the laptop to AC power
5. Wait a few seconds until the computer wouldn't respond. Return to step 4 if the laptop doesn't freeze or turn off after you've waited for a while.

Additional context

Although I haven't tested it yet, I'm sure that this behavior would also occur on battery if RADEON_DPM_PERF_LEVEL_ON_BATT is set to high in the config file.

I understand that this can be considered as a kernel- or firmware-related upstream issue (see https://gitlab.freedesktop.org/drm/amd/-/issues/934), but the weird thing about this is that it only happens specifically with TLP 1.4.0. Therefore, I would love to see a workaround being implemented until that mentioned upstream bug ticket finally gets fixed.

Here are the specs of my laptop and relevant software details:

• Model: Lenovo ThinkPad E595
• CPU: AMD Ryzen 7 3700U
• GPU: AMD Radeon RX Vega 10 iGPU (Picasso generation)
• Distribution: Manjaro
• Kernel versions: 5.14.9, 5.10.70
• linux-firmware version: 20210919.d526e041 (also occurred with 20210818.c46b8c3 installed)
• TLP version: 1.4.0
• DE: KDE Plasma 5.22.5
• Compositor: KWin
• Display Manager: SDDM

Gist links:

• tlp-stat output
• Part of journalctl logs around the time of graphics reset episode

[linrunner]

You're right, I consider this a kernel (or firmware) bug. I do not plan to implement a workaround. TLP would consist of nothing else if I did this every time.