[xdp-hints] Re: [RFC bpf-next v3 09/14] net/mlx5e: Implement devtx kfuncs

[Stanislav Fomichev <sdf@google.com>]

On 07/12, Jakub Kicinski wrote:
> On Wed, 12 Jul 2023 12:42:35 -0700 Alexei Starovoitov wrote:
> > > Basically, add to AF_XDP what we already have for its predecessor
> > > AF_PACKET: setsockopt PACKET_VNET_HDR?
> > >
> > > Possibly with a separate new struct, rather than virtio_net_hdr. As
> > > that has dependencies on other drivers, notably virtio and its
> > > specification process.  
> > 
> > yeah. Forgot about this one.
> > That's a perfect fit. I would reuse virtio_net_hdr as-is.
> > Why reinvent the wheel?
> > It would force uapi, but some might argue it's a good thing.
> 
> I was gonna reply on the other leg of the thread but it sounds like
> we're in agreement now? 👏️  virtio_net_hdr is the kind of generic
> descriptor I had in mind.
> 
> I'd suggest breaking hdr_len into L2len, L3len and L4len, tho. How does
> virtio do IP length updates during TSO if it doesn't know where L3
> starts? HW will want to know, and it's easy to add them together in
> cases where it doesn't. Which is why I kept saying "packet geometry"
> rather than pointing at virtio_net_hdr.

Perfect, I'll drop the kfuncs and will switch to a more static way
of passing this info. We can discuss the particular layout once I have
something concrete to show :-)

Thanks, again, everyone for the comments!