Re: [PATCH net-next 14/22] vxlan: Flow based tunneling

[Thomas Graf <tgraf@suug.ch>]

On 07/21/15 at 10:30am, Alexei Starovoitov wrote:
> RX:
> >+		info->mode = IP_TUNNEL_INFO_RX;
> >+		info->key.tun_flags = TUNNEL_KEY;
> >+		info->key.tun_id = cpu_to_be64(vni >> 8);
> ...
> TX:
> >+		dst_port = info->key.tp_dst ? : vxlan->dst_port;
> >+		vni = be64_to_cpu(info->key.tun_id);
> 
> I think the copy paste of ovs_tunnel_info into ip_tunnel_info
> can be improved. In particular instead of '__be64 tun_id'
> we can use '__u64 tun_id' which will avoid extra byteswaps for rx/tx
> paths.
> 
> netlink for this part also seems inconsistent.
> In the patch 16:
> +static const struct nla_policy ip_tun_policy[IP_TUN_MAX + 1] = {
> +	[IP_TUN_ID]		= { .type = NLA_U64 },
> ...
> +	if (tb[IP_TUN_ID])
> +		tun_info->key.tun_id = nla_get_u64(tb[IP_TUN_ID]);
> 
> I think nla_get_be64 should be there?
> and with my suggestion we can add be64_to_cpu() here instead
> of doing it per packet.
> Thoughts?

I like this. The be64 originates from how OVS stores the tun_id in the
flow key. I agree that it makes sense to limit and delay the byteswaps
to when OVS inherits the flow key from the ip_tunnel_info. I will send
a follow-up.