Virtio

Paravirtualized drivers for kvm/Linux

• Virtio was chosen to be the main platform for IO virtualization in KVM
• The idea behind it is to have a common framework for hypervisors for IO virtualization
• More information (although not uptodate) can be found in kvm AttachFile&do=get&target=kvm_pv_drv.pdf pv driver
• At the moment network/block/balloon devices are suported for kvm
• The host implementation is in userspace - qemu, so no driver is needed in the host.

How to use Virtio

• Get kvm version >= 60
• Get Linux kernel with virtio drivers for the guest
  • Get Kernel >= 2.6.25 and activate (modules should also work, but take care of initramdisk)
    • CONFIG_VIRTIO_PCI=y (Virtualization -> PCI driver for virtio devices)
    • CONFIG_VIRTIO_BALLOON=y (Virtualization -> Virtio balloon driver)
    • CONFIG_VIRTIO_BLK=y (Device Drivers -> Block -> Virtio block driver)
    • CONFIG_VIRTIO_NET=y (Device Drivers -> Network device support -> Virtio network driver)
    • CONFIG_VIRTIO=y (automatically selected)
    • CONFIG_VIRTIO_RING=y (automatically selected)
    • CONFIG_HIGH_RES_TIMERS=y (Processor type and features -> High Resolution Timer Support **optional**)
    • you can safely disable SATA/SCSI and also all other nic drivers if you only use VIRTIO (disk/nic)
• Either build it around Rusty's tree repo
  • Or git clone git://kvm.qumranet.com/home/dor/src/linux-2.6-nv use branch rusty
  • Soon an official repository will be released
  • As an alternative one can use a standard guest kernel for the guest > 2.6.18 and make use sync backward compatibility option
  • Backport and instructions can be found in Anthony Liguori's virtio-ext-modules
  • At the moment it's broken since the guest got developed, soon update
• Use model=virtio for the network devices and if=virtio for disk
  • Example

qemu/x86_64-softmmu/qemu-system-x86_64 -boot c -drive file=/images/xpbase.qcow2,if=virtio,boot=on -m 384 -net nic,model=virtio -net tap,script=/etc/kvm/qemu-ifup

• -hd[ab] for disk won't work, use -drive
• Disk will show up as /dev/vd[a-z][1-9], if you migrate you need to change "root=" in Lilo/GRUB config
• At the moment the kernel modules are automatically loaded in the guest but the interface should be started manually (dhclient/ifconfig)
• Currently performance is much better when using a host kernel configured with CONFIG_HIGH_RES_TIMERS. Another option is use HPET/RTC and -clock= qemu option.
• Expected performance
  • Performance varies from host to host, kernel to kernel
  • On my laptop I measured 1.1Gbps rx throughput using 2.6.23, 850Mbps tx.
  • Ping latency is 300-500 usec
• Enjoy, more to come :)