Enabling GPU usage meter causes audio stuttering in Pipewire

[joepie91]

I realize this issue title is going to sound absurd, and I genuinely have no idea how this is even possible or what parts of the system are involved, but no matter what I look at, this is the conclusion I come to.

I've been trying to debug audio stuttering with Pipewire (caused by xruns) on my desktop for over a year now, trying everything under the sun to try and solve it. I've fixed the realtime setup (seems to work now), disabled the AMD fTPM (seems to have created a minor improvement), updated across two full distro releases by now (no real change), disabled pipewire-pulse (reduced it, didn't fix it), and I've spent entirely too many hours of my life staring at pw-top and various logs.

Historically, the stuttering and skipping has been particularly prominent when there are multiple active Pipewire clients; pavucontrol and (after removing PulseAudio compatibility) pwvucontrol have been notorious 'amplifiers' of the issue, presumably because of the many connections they make, but nothing has reliably reproduced the issue, except one thing: running htop with the GPU meter enabled. Disable GPU meter, and the ERR count in pw-top stabilizes, even with htop still running. Enable it again, and it starts intermittently ticking up again. Not a lot, somewhere between 'once every 5 minutes' and '10 times a minute', but consistently.

For this reproduction I moved out my usual htoprc so it started from a clean slate with no other changes from a default config besides the GPU meter.

I'm... at a loss for where to even start looking. So I hope that you might have an idea on where to look, and I'd be happy to provide more information or try things out if needed to narrow down the issue (though recompiling htop may take a while for me to do, since that's probably slightly less trivial on NixOS, and I've not exactly been blessed with large amounts of focus lately).

My system:

• Motherboard: Gigabyte Aorus X570 Pro
• CPU: AMD Ryzen 9 3900X
• GPU: AMD Radeon 5700 XT
• OS: NixOS 25.11
• Desktop environment: Budgie
• Audio: Pipewire with WirePlumber, using motherboard DAC (ALC1220)
• htop version: 3.4.1

[BenBE]

Can you try recording a userland+kernel perf trace and see where time is spent? Also taking a look at call latency for the syscalls in pipewire and htop may offer some more clues to the true origin of the stuttering.

There have historically been some "known" issues with reading certain information from the kernel (in particular memory map information), that cause quite some kernel delays. Having the GPU meter active might trigger similar delays inside the kernel and thus break the required response times to guarantee stutter-free audio playback.

Also having a comparison between perf traces for htop running with and without the GPU meter might provide some starting point for further investigation.