[BUG] Running vrcompositor-launcher with cap_sys_nice causes graphics reset and system crash when playing VRChat

[Jibodeah]

Describe the bug
When SteamVR is allowed to set cap_sys_nice on vrcompositor launcher during setup (the step in which it asks for admin), a graphics timeout can occur when playing VRChat, rendering the system unusable until restarted. Green graphical artefacts also appear in the headset in VRChat's loading screens.

To Reproduce
Steps to reproduce the behaviour:

1. Install SteamVR, allowing it to set cap_sys_nice during setup.
2. Play VRChat (note green artefacts in the headset when VRChat is loading worlds)
3. Generally within two hours of playtime the crash occurs.

Expected behaviour
Should not crash, and the artefacts should not appear either.

System Information:

• Distribution: Fedora
• SteamVR version: 1.25.7
• Steam client version: 6th April 2023 at 23:02:22
• Opted into Steam client beta?: Yes
• Graphics driver version: Mesa 22.3.7
• Gist for SteamVR System Information: https://gist.github.com/Jibodeah/b99e1bdaaeccc0d7f9caa9bf93e50d61

Additional context
I originally posted about this in a comment on another issue, not knowing then what exactly caused it. I figure now it deserves its own post since it doesn't seem to be exactly the same issue as that one.

Its possible this may be an upstream bug from mesa, or dxvk or such. I don't know enough about this stuff to be sure.

Here's an extract from the syslog from one of these crashes.

Jan 17 01:36:02 apathia kernel: [drm:amdgpu_job_timedout [amdgpu]] *ERROR* ring gfx_0.0.0 timeout, signaled seq=1429119, emitted seq=1429121
Jan 17 01:36:02 apathia kernel: [drm:amdgpu_job_timedout [amdgpu]] *ERROR* Process information: process VRChat.exe pid 9075 thread dxvk-submit pid 9119
Jan 17 01:36:02 apathia kernel: amdgpu 0000:03:00.0: amdgpu: GPU reset begin!
[...]
Jan 17 01:36:18 apathia kernel: [drm:psp_v11_0_memory_training [amdgpu]] *ERROR* send training msg failed.
Jan 17 01:36:18 apathia kernel: [drm:psp_resume [amdgpu]] *ERROR* Failed to process memory training!
Jan 17 01:36:18 apathia kernel: [drm:amdgpu_device_fw_loading [amdgpu]] *ERROR* resume of IP block <psp> failed -62
Jan 17 01:36:18 apathia kernel: [drm] Skip scheduling IBs!
Jan 17 01:36:18 apathia kernel: [drm] Skip scheduling IBs!
Jan 17 01:36:18 apathia kernel: [drm] Skip scheduling IBs!
Jan 17 01:36:18 apathia kernel: amdgpu 0000:03:00.0: amdgpu: GPU reset(2) failed
Jan 17 01:36:18 apathia kernel: [drm] Skip scheduling IBs!
[above line repeats many times]

I have only observed this crash while playing VRChat, and only while playing VRChat in VR. Perhaps other games are affected too.

As noted above, allowing cap_sys_nice to be set also seems to cause green graphical artefacts in the headset during VRChat's loading screens. I don't know if these correlate 100% with the crashes, but they seem to.

Note: Commenters who are also experiencing this issue are encouraged to include the "System Information" section in their replies.

[purplebar0]

The artifacts are always present for the first couple of seconds, and the crashes occur after a couple of minutes (up to an hour) of SteamVR running. Crashes happen in SteamVR Home and any other game I tested, I presume that it happens even if no game is running.

The workaround is to ignore the dialogue that asks for root access, but if the capability has already been set, it should be removed:
sudo setcap -r ~/.steam/steam/steamapps/common/SteamVR/bin/linux64/vrcompositor-launcher

Having the capability set didn't make a big difference in my experience, it's clear that the idea is to give itself a lower niceness value, but I have no idea whether that by itself causes issues, I presume not.

System report