QEMU/KVM architecture overview

it’s a note of “KVM Arch” slide in 2015 KVM forum

Qemu process model (Qemu and Guest OS)

  • linux userspace process: process memory = qemu memory + guest OS physical memory

  • each KVM vCPU is a thread.

    • Host kernel scheduler decides when vCPUs run

Event-driven multi-threaded

  • event loop used for timer, fd, monitoring … etc

    • non-blocking IO

    • callback or coroutines

  • multi-thread architecture but with big lock

    • vcpu threads execute in parallel

    • specific task runs in other threads: RAM live migration, remote displaying encoding, virtio-blk dataplane

    • global mutex

see QEMU thread and event model for detail

host/guest device emulation split

  • guest device: device model visible to guest

  • host device: performs IO on behalf of guest

use QEMU CLI option as example.

  • There are 2 parts to networking in QEMU

    1. virtual network device: -device <device>,netdev=<id>

    2. network backend: -netdev <backend>,id=<id>

  • 2 examples

    # guest device=e1000 & host device=user (user network/slirp)
-netdev user,id=net0 \
-device e1000,netdev=net0

# guest device=virtio-net & host device=user
-netdev user,id=net0 \
-device virtio-net-pci,netdev=net0

virtio devices

vhost in-kernel devices

  • vhost drivers emulate virtio devices in host kernel for better performance

    • vhost_net.ko: high-performance virtio-net emulation, kernel-only zero-copy and interrupt handling features.

Qemu