CPU governor for pinned KVM vCPUs

The action in this article has been conducted under the following conditions:

Property	Value
Date	September 2026
Kernel	6.8.0-139-generic
Distribution	Linux Mint 22.3
QEMU	8.2.2
libvirt	10.0.0

Problem

Pinned guest vCPUs only stay on the host CPUs you assign. They do not get a performance frequency governor by themselves.

A common first step is a udev rule that sets scaling_governor=performance on those CPUs at ACTION=="add" (device appear at boot). That works on a quiet boot. After a Debian / Ubuntu / Linux Mint upgrade, the cpufrequtils LSB service typically starts next and writes ondemand (or whatever GOVERNOR is) to every CPU. The udev change is gone before login.

Confirm with:

cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

If every line is ondemand even though the udev rule exists, this is the race.

This is not CPU isolation (isolcpus) and not systemd CPUAffinity. Those keep other tasks off the guest CPUs. The governor only controls how high those CPUs clock when they do run work.

Approach

Keep the udev rule as a first attempt. Make the setting survive cpufrequtils with a oneshot unit that runs after it.

Host CPUs used by the desktop stay on ondemand (via cpufrequtils).
Host CPUs listed in the guest <cputune> / cpuset are forced to performance.

Example: a guest pinned to host CPUs 8-15 uses that same list here. Substitute the set from virsh vcpuinfo <domain> / the domain XML.

Host files

The unit, helper, and defaults below are also in host/ in this repository. The optional udev rule is not. From a checkout:

sudo sh docs/research/VFIO/host/install.sh 8-15

Without the CPU-list argument the script only installs the sample VM_CPUS=8-15 file and does not enable the unit. Pass this host’s guest pin list to write it and systemctl enable --now. An existing /etc/default/cpufrequtils is left in place.

1. CPU list

/etc/default/vm-cpu-governor

# Host CPUs reserved for pinned KVM guests.
# Use the same set as <vcpu cpuset="..."> / <cputune><vcpupin>.
VM_CPUS=8-15
VM_GOVERNOR=performance

2. Keep cpufrequtils from becoming a silent default

/etc/default/cpufrequtils is optional but useful: the init script sources it if present. This keeps host-wide policy explicit after package upgrades.

ENABLE="true"
GOVERNOR="ondemand"
MAX_SPEED="0"
MIN_SPEED="0"

Leave ENABLE="true" so leftover host CPUs still get ondemand. Do not set GOVERNOR="performance" here unless you want the entire machine on that governor.

3. Helper script

/usr/local/sbin/set-vm-cpu-governor

#!/bin/sh
# Apply a cpufreq governor to the host CPUs reserved for pinned KVM guests.
set -eu

VM_CPUS=8-15
VM_GOVERNOR=performance

if [ -r /etc/default/vm-cpu-governor ]; then
    # shellcheck disable=SC1091
    . /etc/default/vm-cpu-governor
fi

expand_cpus() {
    echo "$1" | tr ',' '\n' | while IFS= read -r tok; do
        tok=$(printf '%s' "$tok" | tr -d '[:space:]')
        [ -z "$tok" ] && continue
        case "$tok" in
        *-*)
            start=${tok%-*}
            end=${tok#*-}
            seq "$start" "$end"
            ;;
        *)
            printf '%s\n' "$tok"
            ;;
        esac
    done
}

status=0
for cpu in $(expand_cpus "$VM_CPUS"); do
    gov="/sys/devices/system/cpu/cpu${cpu}/cpufreq/scaling_governor"
    if [ ! -f "$gov" ]; then
        echo "set-vm-cpu-governor: missing $gov" >&2
        status=1
        continue
    fi
    printf '%s\n' "$VM_GOVERNOR" > "$gov"
done
exit "$status"

sudo chmod 755 /usr/local/sbin/set-vm-cpu-governor

4. systemd oneshot

/etc/systemd/system/vm-cpu-governor.service

[Unit]
Description=Set cpufreq governor on host CPUs reserved for KVM guests
After=cpufrequtils.service
After=loadcpufreq.service
Wants=loadcpufreq.service

[Service]
Type=oneshot
ExecStart=/usr/local/sbin/set-vm-cpu-governor
RemainAfterExit=yes

[Install]
WantedBy=multi-user.target

sudo systemctl daemon-reload
sudo systemctl enable --now vm-cpu-governor.service

5. Optional udev rule (manual only)

This rule is not in host/ and install.sh does not install it. Create the file by hand if you still want the boot-time first attempt. It is not sufficient on its own after cpufrequtils is installed.

/etc/udev/rules.d/90-scaling-governor-performance.rules

KERNEL=="cpu8|cpu9|cpu10|cpu11|cpu12|cpu13|cpu14|cpu15", SUBSYSTEM=="cpu", ACTION=="add", ATTR{cpufreq/scaling_governor}="performance"

Adjust the KERNEL== match if your guest CPU set is not 8-15.

Verify

systemctl status vm-cpu-governor.service --no-pager
for i in $(seq 0 15); do
  printf 'cpu%-2s %s\n' "$i" "$(cat /sys/devices/system/cpu/cpu$i/cpufreq/scaling_governor)"
done

Expected pattern on a 16-thread host with guests pinned to 8-15:

cpu0–cpu7: ondemand
cpu8–cpu15: performance

Re-check after a reboot. If everything is ondemand again, vm-cpu-governor.service is not enabled or After=cpufrequtils.service is missing.

Notes

Changing the governor does not pin QEMU threads. Keep <cputune> in the domain XML.
Keeping other host tasks off those CPUs is a separate step: Host CPU affinity for pinned VMs.
isolcpus= is a more aggressive option than slice AllowedCPUs=. It was left unused here.
On AMD Zen, Linux CPU numbers N and N+core_count are usually SMT siblings of the same physical core. Pin and governor lists should still match whatever you already chose in libvirt; this unit does not change that layout.