XML with qemu.conf

Incus
1

Hello,

I know that we can override qemu configuration using raw.qemu.conf, but I don’t know how to override them using XML.

i would like to override my conf using this snippet to test something but I don’t know how to convert it:

  <clock offset='utc'>
    <timer name='hpet' present='yes'/>
    <timer name='hypervclock' present='yes'/>
  </clock>

If somebody has a tip, thank you

2

QEMU doesn’t understand XML. I’m guessing your XML is in libvirt’s Domain XML. Porting it isn’t trivial, but you may want to have a look at the hpet bool of the machine definition, the hv-time of the cpu definition and other so-called Hyper-V enlightenments. This is outside of my expertise, but that’s the rough mapping I found.

Although I’m wondering why you want that, as we already have clock logic dedicated to Windows guests in our QEMU driver logic.

3

Yes, I’m talking about libvirt.

I want to do that because I have high CPU usage on my host despite the usage in VM is low (30% on host with 5% in the VM).

Some say that it would help (home assistant forum).

The VM are Linux guests

4

Incus doesn’t use libvirt, so you’ll need to convert your libvirt XML to either QEMU’s .conf format or to QMP commands.

5

Yes I noticed that thank for the hint.

Do you have an idea why the qemu process cpu consumption on host is way higher than the one reported in the vm ?

6

I have the same issue. Depending on the image used, when the guest is idle qemu on the host would use 5-20% CPU.

The fix is this: incus config set trixie-vm raw.qemu='-machine hpet=off'

Now CPU is down to 2-4% per VM.

There’s probably a reason why incus doesn’t disable hpet by default, though?

7

I found the second culprit: incus uses virtio-scsi instead of virtio-blk which causes the guest to constantly poll. This is the remaining 2-4% of CPU usage. I managed to force it by patching incus and it did fall to 0% (like proxmox/cockpit and libvirt do) instead of the 2-4% that we get after disabling hpet.

Unfortunately in another thread @stgraber mentioned that incus would stick with scsi because it is better supported by guests.

This is unfortunate but understandable. Though a way to override it would be nice (I couldn’t find one without patching incus itself).

In summary:

Default incus: 20% idle CPU usage

Default incus with raw.qemu override for machine hpet=off: 2-4% idle CPU usage

Patched incus with hpet by default, and virtio-blk: 0% idle CPU usage

8

You can just set io.bus=virtio-blk on your disk device.

9

Thank you that will definitely be useful to me! Also this is embarrassing and I’m sorry for invoking you unnecessarily: I was coming back to stealth edit my previous comment because I was wrong. I had changed two things at once, stupidly.

Removing ‘[device “qemu_gpu”]’ through raw.qemu.conf is what actually brought idle CPU usage down to zero. Libvirt has the exact same device definition of virtio-vga and yet achieves 0% idle so it might be something related to the gpu but not the gpu itself.

SPICE, perhaps? Libvirt uses VNC by default instead. That’s about all I can see. Or perhaps it’s a kernel config difference (in the guest).

I couldn’t find a way to disable SPICE to test, so I will patch it out of incus and try tomorrow. I’ve seen that 7.1 added feature detection for it so possibly I could remove support from my system and incus would simply not try to use it?

Edit:
This is the magic line for nearly 0% idling of guest, it has to be done for each vm container:
incus config set <vm-name> raw.qemu='-machine hpet=off' raw.qemu.conf='[device "qemu_gpu"]'

1 Like