Re: [Qemu-devel] updated: kvm networking todo wiki

[Rusty Russell]

Anthony Liguori <address@hidden> writes:
> "Michael S. Tsirkin" <address@hidden> writes:
>
>> On Thu, May 30, 2013 at 08:40:47AM -0500, Anthony Liguori wrote:
>>> Stefan Hajnoczi <address@hidden> writes:
>>> 
>>> > On Thu, May 30, 2013 at 7:23 AM, Rusty Russell <address@hidden> wrote:
>>> >> Anthony Liguori <address@hidden> writes:
>>> >>> Rusty Russell <address@hidden> writes:
>>> >>>> On Fri, May 24, 2013 at 08:47:58AM -0500, Anthony Liguori wrote:
>>> >>>>> FWIW, I think what's more interesting is using vhost-net as a 
>>> >>>>> networking
>>> >>>>> backend with virtio-net in QEMU being what's guest facing.
>>> >>>>>
>>> >>>>> In theory, this gives you the best of both worlds: QEMU acts as a 
>>> >>>>> first
>>> >>>>> line of defense against a malicious guest while still getting the
>>> >>>>> performance advantages of vhost-net (zero-copy).
>>> >>>>>
>>> >>>> It would be an interesting idea if we didn't already have the vhost
>>> >>>> model where we don't need the userspace bounce.
>>> >>>
>>> >>> The model is very interesting for QEMU because then we can use vhost as
>>> >>> a backend for other types of network adapters (like vmxnet3 or even
>>> >>> e1000).
>>> >>>
>>> >>> It also helps for things like fault tolerance where we need to be able
>>> >>> to control packet flow within QEMU.
>>> >>
>>> >> (CC's reduced, context added, Dmitry Fleytman added for vmxnet3 
>>> >> thoughts).
>>> >>
>>> >> Then I'm really confused as to what this would look like.  A zero copy
>>> >> sendmsg?  We should be able to implement that today.
>>> >>
>>> >> On the receive side, what can we do better than readv?  If we need to
>>> >> return to userspace to tell the guest that we've got a new packet, we
>>> >> don't win on latency.  We might reduce syscall overhead with a
>>> >> multi-dimensional readv to read multiple packets at once?
>>> >
>>> > Sounds like recvmmsg(2).
>>> 
>>> Could we map this to mergable rx buffers though?
>>> 
>>> Regards,
>>> 
>>> Anthony Liguori
>>
>> Yes because we don't have to complete buffers in order.
>
> What I meant though was for GRO, we don't know how large the received
> packet is going to be.  Mergable rx buffers lets us allocate a pool of
> data for all incoming packets instead of allocating max packet size *
> max packets.
>
> recvmmsg expects an array of msghdrs and I presume each needs to be
> given a fixed size.  So this seems incompatible with mergable rx
> buffers.

Good point.  You'd need to build 64k buffers to pass to recvmmsg, then
reuse the parts it didn't touch on the next call.  This limits us to
about a 16th of what we could do with an interface which understood
buffer merging, but I don't know how much that would matter in
practice.  We'd need some benchmarks....

Cheers,
Rusty.