virtio specification: OOO completion of the xmit buffers in the networking device

virtio specification: OOO completion of the xmit buffers in the networking device

[Vlad Zolotarov]

Rusty, hi!
I'd like to ask for a small clarification about the virtio spec.
The virtio specification is put the way that it allows the out-of-order 
completions in general. Although it states that it's mostly relevant to 
the virtio-blk it doesn't explicitly bans it for virtio networking 
devices (for generality?).

(AFAICS the current virtio HV implementations make sure there is no OOO 
Tx buffers completions).

I wonder if this feature (OOO Tx buffers completions) seems practical to 
u and if not can I hope that it could be clearly stated in the future 
virtio specification releases that it's forbidden?

thanks in advance,
vlad

Re: virtio specification: OOO completion of the xmit buffers in the networking device

[Rusty Russell]

Vlad Zolotarov <vladz@cloudius-systems.com> writes:
> Rusty, hi!
> I'd like to ask for a small clarification about the virtio spec.
> The virtio specification is put the way that it allows the out-of-order 
> completions in general. Although it states that it's mostly relevant to 
> the virtio-blk it doesn't explicitly bans it for virtio networking 
> devices (for generality?).

Indeed.

> (AFAICS the current virtio HV implementations make sure there is no OOO 
> Tx buffers completions).
>
> I wonder if this feature (OOO Tx buffers completions) seems practical to 
> u and if not can I hope that it could be clearly stated in the future 
> virtio specification releases that it's forbidden?

Are you sure that vhost-net doesn't allow async completion already?
I can't immediately see why it can't happen, but MST is CC'd.

Cheers,
Rusty.

Re: virtio specification: OOO completion of the xmit buffers in the networking device

[Michael S. Tsirkin]

On Wed, Jun 04, 2014 at 01:22:49PM +0930, Rusty Russell wrote:
> Vlad Zolotarov <vladz@cloudius-systems.com> writes:
> > Rusty, hi!
> > I'd like to ask for a small clarification about the virtio spec.
> > The virtio specification is put the way that it allows the out-of-order 
> > completions in general. Although it states that it's mostly relevant to 
> > the virtio-blk it doesn't explicitly bans it for virtio networking 
> > devices (for generality?).
> 
> Indeed.
> 
> > (AFAICS the current virtio HV implementations make sure there is no OOO 
> > Tx buffers completions).
> >
> > I wonder if this feature (OOO Tx buffers completions) seems practical to 
> > u and if not can I hope that it could be clearly stated in the future 
> > virtio specification releases that it's forbidden?
> 
> Are you sure that vhost-net doesn't allow async completion already?
> I can't immediately see why it can't happen, but MST is CC'd.
> 
> Cheers,
> Rusty.

vhost-net completes in order but this is just an implementation detail:
this allows keeping a couple of indexes for head/tail instead
of tracking requests individually.

We can change this in the future and
guest must allow out of order completions.

-- 
MST