Virtualization - Dec 2020 - [PATCH rfc 1/3] virtio-net: support transmit hash report

On Mon, Dec 28, 2020 at 11:28 AM Michael S. Tsirkin <mst at redhat.com>
wrote:
>
> On Mon, Dec 28, 2020 at 11:22:31AM -0500, Willem de Bruijn wrote:
> > From: Willem de Bruijn <willemb at google.com>
> >
> > Virtio-net supports sharing the flow hash from host to guest on rx.
> > Do the same on transmit, to allow the host to infer connection state
> > for more robust routing and telemetry.
> >
> > Linux derives ipv6 flowlabel and ECMP multipath from sk->sk_txhash,
> > and updates these fields on error with sk_rethink_txhash. This feature
> > allows the host to make similar decisions.
> >
> > Besides the raw hash, optionally also convey connection state for
> > this hash. Specifically, the hash rotates on transmit timeout. To
> > avoid having to keep a stateful table in the host to detect flow
> > changes, explicitly notify when a hash changed due to timeout.
>
> I don't actually see code using VIRTIO_NET_HASH_STATE_TIMEOUT_BIT
> in this series. Want to split out that part to a separate patch?
Will do.

I wanted to make it clear that these bits must be reserved (i.e.,
zero) until a later patch specifies them.

The timeout notification feature requires additional plumbing between
the TCP protocol stack and device driver, probably an skb bit. I'd
like to leave that as follow-up for now.

Thanks for the fast feedback!

On Mon, Dec 28, 2020 at 11:47:45AM -0500, Willem de Bruijn
wrote:
> On Mon, Dec 28, 2020 at 11:28 AM Michael S. Tsirkin <mst at
redhat.com> wrote:
> >
> > On Mon, Dec 28, 2020 at 11:22:31AM -0500, Willem de Bruijn wrote:
> > > From: Willem de Bruijn <willemb at google.com>
> > >
> > > Virtio-net supports sharing the flow hash from host to guest on
rx.
> > > Do the same on transmit, to allow the host to infer connection
state
> > > for more robust routing and telemetry.
> > >
> > > Linux derives ipv6 flowlabel and ECMP multipath from
sk->sk_txhash,
> > > and updates these fields on error with sk_rethink_txhash. This
feature
> > > allows the host to make similar decisions.
> > >
> > > Besides the raw hash, optionally also convey connection state for
> > > this hash. Specifically, the hash rotates on transmit timeout. To
> > > avoid having to keep a stateful table in the host to detect flow
> > > changes, explicitly notify when a hash changed due to timeout.
> >
> > I don't actually see code using VIRTIO_NET_HASH_STATE_TIMEOUT_BIT
> > in this series. Want to split out that part to a separate patch?
> 
> Will do.
> 
> I wanted to make it clear that these bits must be reserved (i.e.,
> zero) until a later patch specifies them.
Already the case for the padding field I think ...
> The timeout notification feature requires additional plumbing between
> the TCP protocol stack and device driver, probably an skb bit. I'd
> like to leave that as follow-up for now.
> 
> Thanks for the fast feedback!