Re: [Qemu-devel] [RFC] Architecture to connect a userspace ethernet switch to QEMU guests via Virtio

[Stefan Hajnoczi <stefanha@gmail.com>]

On Thu, Nov 21, 2013 at 12:01:20PM +0100, Antonios Motakis wrote:
> On Tue, Nov 19, 2013 at 5:11 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> > On Tue, Nov 19, 2013 at 11:17:40AM +0100, Antonios Motakis wrote:
> >> There have been discussions before on these lists on the topic of
> >> connecting a QEMU guest running a virtio_net driver, with an external
> >> userspace ethernet switch (Snabbswitch in particular). The essential
> >> requirement in this is to put the virtio backend in the external
> >> userspace process.
> >>
> >> The preferred direction should be similar to vhost, with the main
> >> difference of the control mechanism being a unix domain socket instead
> >> of an ioctl interface, and of course placing the backend in an
> >> userspace process instead the kernel.
> >>
> >> Since we are pursuing this direction,we would like to share a more
> >> detailed description of the architecture we are working on. Any
> >> feedback is most welcome. It is available here:
> >> http://www.virtualopensystems.com/media/snabbswitch/rfc_snabbswitch_qemu.pdf
> >
> > It sounds like you are proposing an interprocess virtio device
> > interface.  QEMU's virtio-pci emulation calls into a new vapp virtio
> > device inside QEMU, which then forwards virtio device calls to the vapp
> > process.
> >
> > This is pretty different from vhost.  vhost only puts rx/tx handling in
> > the kernel.  Other functionality is handled by plain old QEMU virtio
> > device emulation (e.g. virtio-net ctrl virtqueue).
> >
> 
> Actually, our intended approach is not that different; it is still
> under the control of the vapp 'client' (QEMU) what virtqueues it will
> pass to the userspace process. There is nothing stopping us from doing
> the same thing and handle the ctrl virtqueue in virtio_net.
> 
> In fact, we are currently evaluating the possibility to implement vapp
> not as a completely new component, but as a feature of vhost. We need
> to decouple some things, however it looks like it is the cleanest
> approach. This makes sense, because virito_net already knows how to
> 'coordinate' with vhost.
> 
> To do this we need to decouple ioctl calls in vhost and add support
> for our (very similar) unix domain socket interface. Also, another
> main difference is that we do not need to set a TAP device (we are
> network backend agnostic, since we let the target process decide what
> to do with network data).
> 
> > In the past device plugin interfaces have been rejected by the community
> > because they can lead to lower code quality (out-of-tree devices) and an
> > avenue to bypass the software license.
> >
> > So what's the alternative?  Reuse as much of the vhost approach as
> > possible and define a userspace network I/O interface instead of a
> > device plugin interface.
> 
> Point taken. It is one of our goals to have the code eventually
> upstreamed., so I hope with the above clarifications, our intended
> solution is deemed more acceptable.

To me it sounds okay to follow a vhost-like interface.  But I think that
Michael Tsirkin and Anthony Liguori's thoughts would be interesting too.

I've CCed them.

Stefan