Re: [PATCH v2] virtio_net: support inner header hash for GRE-encapsulated packets

[Jason Wang <jasowang@redhat.com>]

å 2022/11/28 10:31, Xuan Zhuo åé:
> On Fri, 25 Nov 2022 06:49:42 -0500, "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > On Fri, Nov 25, 2022 at 12:16:05PM +0800, Jason Wang wrote:
> > > On Tue, Nov 22, 2022 at 5:08 PM Heng Qi <hengqi@linux.alibaba.com> wrote:
> > > > When VIRTIO_NET_F_RSS is negotiated and the tunnel is used to
> > > > encapsulate the packets, the hash calculated using the outer header
> > > > of the receive packets is always fixed for the same flow packets,
> > > > i.e. they will be steered to the same receive queue.
> > > >
> > > > We add a VIRTIO_NET_HASH_TYPE_GRE_INNER bitmask in \field{hash_types},
> > > > which instructs the device to calculate the hash using the inner
> > > > headers of GRE-encapsulated packets, and a VIRTIO_NET_HASH_REPORT_GRE
> > > > value in \field{hash_tunnel} to report packet type when calculating
> > > > hash over the inner header.
> > > So I think we need a new feature bit for this to keep migration compatibility.
> > I am not sure I agree.  hash types need to be specified or migrated.
> I think migration is necessary to consider. If VM migrates, it may cause VM's
> behavior to be inconsistent.


Yes, this is a behavior that could be detected by the driver: consider 
GRE is supported on src but not dst, we can't live migrate in this case 
without a feature bit.


> > Also having feature bits is a lot of work and duplication,
> > and inconsistency - we do not have bits for existing hash types.
> The existing Hash Types are brought by default when negotiating RSS. I don't
> think it is necessary to negotiate for them separately.
>
> As far as Virtio-Net's Feature Bits are concerned, I am indeed worried, because
> we think Virtio-Net can also add a lot of new features. In the end, there will
> be no feature bit available.


That's the price we need to pay for supporting live migration. Actually, 
transport layer has sufficient bits for us to use, so it's not a big deal.

One thing that might help is to try to reduce the bit usage, e.g 
introduce a _F_TUNNEL_HASH bit for vxlan, ipip etc as well as GRE.

Thanks