[linux[ Visual bugs with Nouveau on Nvidia 2070 (Turing generation)

[TechniKris]

Type

Cosmetic

Bug description

There are multiple visual bugs, all seemingly related to circular, pie chart-shaped objects.
It seems like Mesa/Nouveau driver might be at fault.
I thought about reporting it upstream at https://gitlab.freedesktop.org/mesa/mesa but I wanted to consult here first if you'd like to diagnose it more in-depth and handle it on your own.

Screenshots or videos

Volume dials:

LegacySongProgress skin element:

2024-10-09.21-43-05-00.02.06.788-00.02.45.208-seg2-reencode.mp4

The "hold for menu" button:

2024-10-09.21-43-05-00.02.56.990-00.03.14.037-seg3.mp4

Grade pie chart on the results screen:

2 cases (that I've found) of specific sliders on beatmaps (note the line going through the middle of the slider):

• beatmap ID: 1264503 osu://edit/00:41:982 (2)

2024-10-09.21-43-05-00.04.16.940-00.04.38.141-seg5.mp4

• beatmap ID: 64896 osu://edit/02:08:082 (1) (argon and argon pro are not affected here)

2024-10-09.21-43-05-00.04.55.469-00.05.24.792-seg6.mp4

• I'm pretty sure I stumbled on such a broken slider at least once more, but I don't remember the beatmap.

Beatmap preview progress bar:

2024-10-14.21-58-11-00.00.01.009-00.00.12.927.mp4

Affected drivers/hardware

I've performed all of the below testing with the default GL / OpenGL renderer, but as a sanity check I tested Deferred / OpenGL and... it seems to be affected as well.

The issue occurs with the Nouveau driver - the Linux open source OpenGL driver for Nvidia GPUs, which is used by the system by default, if there's no Nvidia official driver installed.
Only 1 out of 3 tested Nvidia GPUs seems to be affected:

• RTX 2070 ❗ affected
• GTX 650 ➖ not affected
• 9300M GS (tested in a different PC - laptop) ➖ not affected

Additionally:

• Nvidia's official drivers (470.256.02 (for 650 and 2070), 560.35.03 (for 2070), 560.35.03 open variant (for 2070)): ➖ not affected
• AMD GPU - Ryzen 5700G's integrated graphics ➖ not affected
• Zink - using equivalent Vulkan drivers (the game was still using GL / OpenGL renderer), on all of the above GPUs (except 9300M GS and GTX 650, which have no- or inadequate Vulkan) ➖ not affected

Since only the RTX 2070 (Turing generation) is affected, I suspect the bug could be specifically in some kind of codepath for GSP-enabled cards, which is Turing and newer, and based on that I'd presume newer cards might also be affected. I don't have any more of such cards though, so I can't verify this.

Workarounds

Zink seems like an adequate solution here, providing good performance on top of NVK (or any other Vulkan driver). Zink can be enabled by setting these environment variables: __GLX_VENDOR_LIBRARY_NAME=mesa MESA_LOADER_DRIVER_OVERRIDE=zink GALLIUM_DRIVER=zink.

Version

I noticed these issues over the course of last couple months, at least as early as 2024.521.2
Tested in-depth and recorded examples on 2024.1009.1

Logs

compressed-logs.zip
I can provide apitrace if needed.

[hwsmm]

You already isolated the problem and concluded that RTX2060 with nouveau GSP is the only affected case here..? There may be some applicable workarounds, but you should consult with Mesa as you mentioned at first since nouveau GSP support is still very early.

[TechniKris]

I have tried explicitly setting nouveau.config=NvGspRm=0 kernel parameter (available to 16/20 and 30 series) to disable anything related to the GSP, but it didn't seem to make any difference in behavior (only in performance, which I think would align with the feature matrix).
I guess this discussion can at least serve as a reference for anyone else stumbling upon the issue. Listing affected cards would help confirm my assumption about the range of GPUs affected.