On 03/21, Willem de Bruijn wrote: > On Thu, Mar 21, 2019 at 12:45 AM Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote: > > > > > > > > On 03/20/2019 08:39 PM, Alexei Starovoitov wrote: > > > > > I think you need to convince Dave and Eric that > > > above surgery is necessary to do the hack in patch 6 with > > > +static DEFINE_PER_CPU(struct sk_buff, bpf_flow_skb); > > > > > > > Yes, this is a huge code churn. > > It touches a lot of lines. But it deduplicates logic that might become > out of sync if we don't do that now. Yeah, I remove a bunch of copy-paste with this change. > > Honestly I believe we are going too far in this series. > > > > > I think the better option it to introduce new prog type that works > > > without skb. I think it can be pretty close to shape and form to xdp. > > > > > That's an interesting alternative. The existing sample dissector > does not access any skb fields aside from vlan. The interface could > be superseded with one that has a more constrained context, like xdp. > > When parsing for headlen in the device driver rx path, a program could > conceivably in the same pass also compute an rxhash in case the > device did not supply an L4 one. And perhaps some hash (the same?) > to speed up GRO flow lookup. This would require a few extra fields in > bpf_flow_keys. With the xdp-like interface, what would be the end goal? Two different program types, one triggering from eth_get_headlen, the other from existing __skb_flow_dissect? Or we can use xdp-like interface everywhere and just pass the first fragment in __skb_flow_dissect (that should be enough, in theory, to get everything we want)? If we end up with two interface, that can be confusing to the users. Which one to use? Why? Why write essentially the same program twice? If we can agree that we switch everything to xpd-like, do we deprecate the skb-one?
