Re: [PATCH RFC 1/3] xdp: Infrastructure to generalize XDP

On 09/21/2016 08:39 AM, Jesper Dangaard Brouer wrote:

On Tue, 20 Sep 2016 17:01:39 -0700 Alexei Starovoitov [off-list ref] wrote:

quoted

  >  - Provides a more structured environment that is extensible to new
  >    features while being mostly transparent to the drivers

don't see that in these patches either.
Things like packet size change (that we're working on) still
has to be implemented for every driver.
Existing XDP_TX, XDP_DROP have to be implemented per driver as well.

Also introduction of xdp.h breaks existing UAPI.
That's not acceptable either.

This response piss me off!@@#$

We are the early stages of XDP development. Users cannot consider XDP a
stable UAPI yet.  I added a big fat warning to the docs here[1].

If you already consider this a stable API, then I will suggest that we
disable XDP or rip the hole thing out again!!! Create a separate tree
where we can cooperate on getting this right, before forcing this
upstream.  I have raise concern about this several times upstream, but
the patches got applied anyway, because you Alexei, promised this was
super extendable and we could still change the APIs later. Maybe you
tricked me?! I've started to look at the details, and I'm not happy with
the extensibility.

I don't quite follow you here, Jesper. If you're talking about uapi-related
extensibility wrt struct xdp_md, then it's done the same way as done for tc
with 'shadow' struct __sk_buff. The concept of having this internal BPF insn
rewriter is working quite well for this, and it is extendable with new meta
data. Wrt return codes we're flexible to add new ones once agreed upon. The
whole XDP config is done via netlink, nested in IFLA_XDP container, so it can
be extended in future with other attrs, flags, etc, for setup and dumping.
The 'breakage' that was mentioned here was related to moving things into xdp.h.
It can be done and we did it similarly in the past with bpf_common.h, but
then really also bpf.h needs to include the new xdp.h header.