On Thu, Dec 07, 2017 at 04:31:27PM +0800, Jason Wang wrote:
>
>
> On 2017年12月07日 16:04, Stefan Hajnoczi wrote:
> > On Thu, Dec 7, 2017 at 7:54 AM, Avi Cohen (A) <> wrote:
> > > There is already a virtio mechanism in which 2 VMs assigned a virtio device , are communicating via a veth pair in the host .
> > > KVM just passes a pointer of the page of the writer VM to the reader VM - resulting in excellent performance (no vSwitch in the middle)
> > > **Question**: What is the advantage of vhost-pci compared to this ?
> > Which mechanism do you mean?
> >
> > vhost-pci will allow VM-to-VM communication without vmexits when
> > polling mode is used. Does the mechanism you are thinking about
> > require vmexits?
> >
> > Stefan
>
> I guess what Avi means is probably veth tap support (RFC) here:
>
> https://www.spinics.net/lists/netdev/msg454040.html
>
> But in fact, we don't need veth at all, by using rx handler trick, we can
> easily implement pair mode on TUN/TAP.
Okay, that is a different use case:
1. VM<->physical NIC: macvtap
2. VM<->host userspace: vhost-user
3. VM<->host/container: vethtap or pair mode
4. VM<->VM: vhost-pci
Stefan
