Note: If you're looking for solutions on how to work with the new AMD 79x0X3D models, see this issue.

vfio-isolate

vfio-isolate is a command line tool for Linux, which aims to facilitate CPU and memory isolation for running virtual machines with guaranteed latency.

Usage: vfio-isolate [OPTIONS] COMMAND1 [ARGS]... [COMMAND2 [ARGS]...]...

Options:
  -v, --verbose                enable verbose output
  -d, --debug                  enable debug output
  -u, --undo-file <undo-file>  Create a file that describes the operations
                               needed to undo

  --help                       Show this message and exit.

Commands:
  compact-memory  compact memory
  cpu-governor    set the CPU governor for the given CPUs
  cpuset-create   create a cpuset
  cpuset-delete   delete a cpuset
  cpuset-modify   modify a cpuset
  drop-caches     drop caches
  irq-affinity    manipulate the IRQ affinity
  move-tasks      move tasks between cpusets
  restore         restore a previous state using an undo file

Usage

To "partion" your CPU between the host and your VM, a mechanism of the Linx kernel, named "cgroups" is used. There exist 2 versions of cgroup, v1 and v2. Some options are only available for cgroups v1.

To find out what version your system is using use the following command

mount | grep cgroup

it will show type cgroup for v1, and cgroup2 vor v2.

CPU sets are part of cgroups, and define the subset of cores processes in a cgroup are allowed to be scheduled on.

For cgroups v1, they also have some other properties:

Feature	Description
cpu-exclusivity	prevents other sibling cpusets to use the cpus of this cpuset
mem-exclusivity	prevents other sibling cpusets to use the memory (NUMA) of this cpuset
mem-migration	only effective on NUMA systems: when enabled, processes in this cpuset will have their memory migrated to the node they are running on
scheduler-load-balance	when enabled, the scheduler will try to load balance processes within the available cpus

For more information, see the kernel documentation at https://www.kernel.org/doc/Documentation/cgroup-v1/cpusets.txt

Non NUMA example (cgroups v2)

In this example, an AMD 5950x is partitioned between host and VM. The example further assumes that you are using systemd, which has created the cgroups system.slice and user.slice already.

The VM is configured so that it exclusively uses the 8 cores of the second die. The last core of the first die is used for emulation and IO work, and the first seven are for the host.

The command to use would be this:

 sudo vfio-isolate \ 
    cpuset-modify --cpus C0-6,16-22 /system.slice \
    cpuset-modify --cpus C0-6,16-22 /user.slice

This will instruct the two existing cgroups that systemd created to only use the first 7 cores. To pin the VM cores, use libvirt:

  <vcpu placement='static'>16</vcpu>
  <iothreads>1</iothreads>
  <cputune>
    <vcpupin vcpu='0' cpuset='8'/>
    <vcpupin vcpu='1' cpuset='24'/>
    <vcpupin vcpu='2' cpuset='9'/>
    <vcpupin vcpu='3' cpuset='25'/>
    <vcpupin vcpu='4' cpuset='10'/>
    <vcpupin vcpu='5' cpuset='26'/>
    <vcpupin vcpu='6' cpuset='11'/>
    <vcpupin vcpu='7' cpuset='27'/>
    <vcpupin vcpu='8' cpuset='12'/>
    <vcpupin vcpu='9' cpuset='28'/>
    <vcpupin vcpu='10' cpuset='13'/>
    <vcpupin vcpu='11' cpuset='29'/>
    <vcpupin vcpu='12' cpuset='14'/>
    <vcpupin vcpu='13' cpuset='30'/>
    <vcpupin vcpu='14' cpuset='15'/>
    <vcpupin vcpu='15' cpuset='31'/>
    <emulatorpin cpuset='7,23'/>
    <iothreadpin iothread='1' cpuset='7,23'/>
    <vcpusched vcpus='0' scheduler='rr' priority='1'/>
    <vcpusched vcpus='1' scheduler='rr' priority='1'/>
    <vcpusched vcpus='2' scheduler='rr' priority='1'/>
    <vcpusched vcpus='3' scheduler='rr' priority='1'/>
    <vcpusched vcpus='4' scheduler='rr' priority='1'/>
    <vcpusched vcpus='5' scheduler='rr' priority='1'/>
    <vcpusched vcpus='6' scheduler='rr' priority='1'/>
    <vcpusched vcpus='7' scheduler='rr' priority='1'/>
    <vcpusched vcpus='8' scheduler='rr' priority='1'/>
    <vcpusched vcpus='9' scheduler='rr' priority='1'/>
    <vcpusched vcpus='10' scheduler='rr' priority='1'/>
    <vcpusched vcpus='11' scheduler='rr' priority='1'/>
    <vcpusched vcpus='12' scheduler='rr' priority='1'/>
    <vcpusched vcpus='13' scheduler='rr' priority='1'/>
    <vcpusched vcpus='14' scheduler='rr' priority='1'/>
    <vcpusched vcpus='15' scheduler='rr' priority='1'/>
    <iothreadsched iothreads='1' scheduler='fifo' priority='98'/>
  </cputune>

Non NUMA example (cgroups v1)

In this example, we have a 6 core, 12 thread CPU from Intel, and we want to leave the first 2 cores for host, IO and emulation work, while giving the remaining 4 cores to the VM.

The command to use would be this:

 # vfio-isolate \ 
    cpuset-create --cpus C0-1,6-7 /host.slice \
    move-tasks / /host.slice

This will move all the processes from the root cgroup to your newly created host.slice, and assigns only the first two physical cores for execution.

To now make your VM use the now remaining idle cores, you can use libvirt:

<vcpu placement='static'>8</vcpu>
  <vcpupin vcpu="0" cpuset="2"/>
  <vcpupin vcpu="1" cpuset="8"/>
  <vcpupin vcpu="2" cpuset="3"/>
  <vcpupin vcpu="3" cpuset="9"/>
  <vcpupin vcpu="4" cpuset="4"/>
  <vcpupin vcpu="5" cpuset="10"/>
  <vcpupin vcpu="6" cpuset="5"/>
  <vcpupin vcpu="7" cpuset="11"/>
  <emulatorpin cpuset="0-1,6-7"/>
  <iothreadpin iothread="1" cpuset="0-1,6-7"/>
</cputune>

To manually undo the previous command:

 sudo vfio-isolate \ 
    cpuset-delete /host.slice

Or you could use the undo feature built into vfio-isolate (see below).

All processes in a cpuset will be moved to its parent cpuset upon deletion.

NUMA example (cgroups v1)

If you have a system with more than one NUMA nodes, you might want to isolate according to the different nodes. For example, on an AMD Threadripper 1920X (12 core, 24 thread), which has 2 NUMA nodes, you could do the following

 # vfio-isolate \ 
    cpuset-create --cpus N0 --mems N0 -mm /host.slice \
    move-tasks / /host.slice

This will configure NUMA Node 0, in this case CPU 0-5,12-17 for the host, while configuring NUMA node 1 for the VM (6-11,18-23). The -mm parameter enables memory migration, so that processes moving into either the host or the VM cpuset will have their memory migrated to the right node.

Undo

vfio-isolate is able to record all the changes that it did and storing a recipe to undo them into a file, to be executed later.

 # vfio-isolate -u /tmp/undo_description \ 
    cpuset-create --cpus C1-4 /test.slice

This will create the test.slice cpuset, and also a file /tmp/undo_description that when executed like this

 # vfio-isolate restore /tmp/undo_description

will remove test.slice. This works with all the subcommands that vfio-isolate supports.

IRQ affinity

vfio-isolate contains basic support for disabling IRQ handler execution on certain cpus:

 # vfio-isolate -u /tmp/undo_irq irq-affinity mask C2-5,8-11

will prevent IRQ execution on the mentioned cpuset. It will also write an undo description in /tmp/undo_irq which can be used to restore the previous state:

 # vfio-isolate restore /tmp/undo_irq

setting CPU governor

vfio-isolate contains basic support for setting the CPU frequency governor for selected CPUs:

 # vfio-isolate -u /tmp/undo_gov cpu-governor performance C2-5,8-11

will set the mentioned CPUs to performance mode. It will also write an undo description in /tmp/undo_gov which can be used to restore the previous state:

 # vfio-isolate restore /tmp/undo_gov

Fully featured example (/etc/libvirt/hooks/qemu):

#!/bin/bash

HCPUS=0-6,16-22
MCPUS=8-15,24-31

UNDOFILE=/var/run/libvirt/qemu/vfio-isolate-undo.bin

disable_isolation () {
	vfio-isolate \
		restore $UNDOFILE

	taskset -pc 0-31 2  # kthreadd reset
}

enable_isolation () {
	vfio-isolate \
		-u $UNDOFILE \
		drop-caches \
		cpu-governor performance \
		cpuset-modify --cpus C$HCPUS /system.slice \
		cpuset-modify --cpus C$HCPUS /user.slice \
		compact-memory \
		irq-affinity mask C$MCPUS

	taskset -pc $HCPUS 2  # kthreadd only on host cores
}

case "$2" in
"prepare")
	enable_isolation
	;;
"started")
	;;
"release")
	disable_isolation
	;;
esac