On Sat, Jan 12, 2013 at 10:50:30AM +1030, Rusty Russell wrote:

"Michael S. Tsirkin" [off-list ref] writes:

quoted

On Fri, Jan 11, 2013 at 09:18:33AM +1030, Rusty Russell wrote:

quoted

"Michael S. Tsirkin" [off-list ref] writes:

quoted

On Thu, Jan 10, 2013 at 09:00:55PM +1030, Rusty Russell wrote:

quoted

Not sure why vhost/net doesn't built a packet and feed it in
netif_rx_ni().  This is what tun seems to do, and with this code it
should be fairly optimal.

Because we want to use NAPI.

Not quite what I was asking; it was more a question of why we're using a
raw socket, when we trivially have a complete skb already which we
should be able to feed to Linux like any network packet.

Oh for some reason I thought you were talking about virtio.
I don't really understand what you are saying here - vhost
actually calls out to tun to build and submit the skb.

Ah, the fd is tun?

It can be tun or macvtap. We also support a packet socket
backend though I don't know of any users, maybe this can
be dropped.

 Seems a bit indirect; I wonder if there's room for
more optimization here...

Cheers,
Rusty.

Quite possibly. Using common data structures and code in tun and macvtap
would allow calling this code directly from vhost-net.