[bisected] [regression] vblank_disable_immediate causes heavy stuttering with 6700XT on Xorg

The Problem

These small hangs last about half a second and can happen over and over multiple minutes (or even seconds) apart making desktop usage extremely annoying.

The kernel prints [drm:dc_dmub_srv_wait_idle [amdgpu]] *ERROR* Error waiting for DMUB idle: status=3 after every stutter.

I've seen the same error message in other reports with RDNA2 cards but none of them seems to describe the same symptoms so this error message is either extremely generic or this bug is even worse than I thought and it's affecting multiple cards in the Navi 2X line.

The Cause

This commit.

Reverting it fixes the problem completely.

I suggest disabling vblank_disable_immediate until a proper solution is found since even light desktop usage is simply too annoying to be viable with this card on newer kernels when using Xorg.

How to Reproduce

Use Xorg and Linux 5.15.17+ (I've been living on 5.15.16 for a while to avoid this bug).

Simply leaving the system idle will eventually trigger this (the screen doesn't have to turn off and the system doens't have to suspend). Simple web browsing with Firefox or reproducing a video seems to trigger it faster.

This can take multiple minutes though (it took me a very long time to bisect this).

I've been able to reproduce this on KDE Plasma (Kwin) and Openbox (with Picom as a compositor).

I could not reproduce this bug on Plasma Wayland at this point but I can't exclude the possibility that other Wayland compositors can be affected.

Hardware Information

GPU: Radeon RX 6700XT

Display connection: DisplayPort

System Information

Distro: Ubuntu 22.04, Ubuntu 20.04

Kernel: 5.15.17+, 5.17+ git

Xorg 1.21.1.3, 1.20.13

Mesa 22.0.1 (radeonsi)

Window managers and compositors: KWin, Openbox+Picom

added DC label

Can confirm I've also seen this issue on my 6700 XT. In my case no stuttering occurred on the desktop (GNOME on X.org) but only when playing games with V-sync enabled in fullscreen. Reverting the commit also fixes the problem for me.

Just in case this helps anyone, I've been able to work around this bug by booting with the drm.vblankoffdelay=0 kernel parameter.

This comes at a power consumption cost though and so this also achieves the exact opposite of what the offending commit was supposed to do (saving power) so that's far from ideal but at least it doesn't require a kernel recompilation.

In addition I can confirm that this also happens when playing games with vsync on. I never noticed it before since I almost always play with vsync off on my freesync monitor.

mentioned in issue #1887

I'm currently using the 5.15 LTS kernel (currently at 5.15.48) and have experienced the same message in my journal ever since I had to take one of my (two) monitors to the shop for inspection/replacement... and here is what I observed with my 6800XT:

• With two monitors connected the lowest GPU power draw is 34W and no sign of the spam message; all was good
• With only one monitor connected the lowest GPU power draw was 8W and the spam message occurred intermittently

So in my experience, keeping two monitors connected kept the GPU at a high enough power consumption to prevent the message from being written to the journal.

With only one monitor connected, I could always tell when the issue arose and the message was logged in the journal while playing a game... because suddenly I would experience stuttering graphics for a few moments before it recovered.

The revert that fixed this bug was merged upstream in this commit. Closing.

closed