On Fri, Nov 27, 2020 at 02:32:44PM +0200, Eyal Birger wrote: > Hi Steffen, > > On Fri, Nov 27, 2020 at 11:44 AM Steffen Klassert > <steffen.klass...@secunet.com> wrote: > > > > On Sat, Nov 21, 2020 at 04:28:23PM +0200, Eyal Birger wrote: > > > This commit adds support for 'collect_md' mode on xfrm interfaces. > > > > > > Each net can have one collect_md device, created by providing the > > > IFLA_XFRM_COLLECT_METADATA flag at creation. This device cannot be > > > altered and has no if_id or link device attributes. > > > > > > On transmit to this device, the if_id is fetched from the attached dst > > > metadata on the skb. The dst metadata type used is METADATA_IP_TUNNEL > > > since the only needed property is the if_id stored in the tun_id member > > > of the ip_tunnel_info->key. > > > > Can we please have a separate metadata type for xfrm interfaces? > > > > Sharing such structures turned already out to be a bad idea > > on vti interfaces, let's try to avoid that misstake with > > xfrm interfaces. > > My initial thought was to do that, but it looks like most of the constructs > surrounding this facility - tc, nft, ovs, ebpf, ip routing - are built around > struct ip_tunnel_info and don't regard other possible metadata types.
That is likely because most objects that have a collect_md mode are tunnels. We have already a second metadata type, and I don't see why we can't have a third one. Maybe we can create something more generic so that it can have other users too. > For xfrm interfaces, the only metadata used is the if_id, which is stored > in the metadata tun_id, so I think other than naming consideration, the use > of struct ip_tunnel_info does not imply tunneling and does not limit the > use of xfrmi to a specific mode of operation. I agree that this can work, but it is a first step into a wrong direction. Using a __be64 field of a completely unrelated structure as an u32 if_id is bad style IMO. > On the other hand, adding a new metadata type would require changing all > other places to regard the new metadata type, with a large number of > userspace visible changes. I admit that this might have some disadvantages too, but I'm not convinced that this justifies the 'ip_tunnel_info' hack.
