Re: [PATCH bpf 1/2] flow_dissector: support FLOW_DISSECTOR_KEY_ETH_ADDRS with BPF

[Stanislav Fomichev <sdf@google.com>]

> On 05/13, Willem de Bruijn wrote:
> > On Mon, May 13, 2019 at 5:21 PM Willem de Bruijn
> > <willemdebruijn.kernel@gmail.com> wrote:
> > >
> > > On Mon, May 13, 2019 at 5:02 PM Stanislav Fomichev <sdf@fomichev.me> wrote:
> > > >
> > > > On 05/13, Willem de Bruijn wrote:
> > > > > On Mon, May 13, 2019 at 3:53 PM Stanislav Fomichev <sdf@google.com> wrote:
> > > > > >
> > > > > > If we have a flow dissector BPF program attached to the namespace,
> > > > > > FLOW_DISSECTOR_KEY_ETH_ADDRS won't trigger because we exit early.
> > > > >
> > > > > I suppose that this is true for a variety of keys? For instance, also
> > > > > FLOW_DISSECTOR_KEY_IPV4_ADDRS.
> > >
> > > > I though the intent was to support most of the basic stuff (eth/ip/tcp/udp)
> > > > without any esoteric protocols.
> > >
> > > Indeed. But this applies both to protocols and the feature set. Both
> > > are more limited.
> > >
> > > > Not sure about FLOW_DISSECTOR_KEY_IPV4_ADDRS,
> > > > looks like we support that (except FLOW_DISSECTOR_KEY_TIPC part).
> > >
> > > Ah, I chose a bad example then.
> > >
> > > > > We originally intended BPF flow dissection for all paths except
> > > > > tc_flower. As that catches all the vulnerable cases on the ingress
> > > > > path on the one hand and it is infeasible to support all the
> > > > > flower features, now and future. I think that is the real fix.
> > >
> > > > Sorry, didn't get what you meant by the real fix.
> > > > Don't care about tc_flower? Just support a minimal set of features
> > > > needed by selftests?
> > >
> > > I do mean exclude BPF flow dissector (only) for tc_flower, as we
> > > cannot guarantee that the BPF program can fully implement the
> > > requested feature.
> >
> > Though, the user inserting the BPF flow dissector is the same as the
> > user inserting the flower program, the (per netns) admin. So arguably
> > is aware of the constraints incurred by BPF flow dissection. And else
> > can still detect when a feature is not supported from the (lack of)
> > output, as in this case of Ethernet address. I don't think we want to
> > mix BPF and non-BPF flow dissection though. That subverts the safety
> > argument of switching to BPF for flow dissection.
> Yes, we cannot completely avoid tc_flower because we use it to do
> the end-to-end testing. That's why I was trying to make sure "basic"
> stuff works (it might feel confusing that tc_flower {src,dst}_mac
> stop working with a bpf program installed).
>
> TBH, I'd not call this particular piece of code that exports src/dst
> addresses a dissection. At this point, it's a well-formed skb with
> a proper l2 header and we just copy the addresses out. It's probably
> part of the reason the original patch didn't include any skb->protocol
> checks.
On the other hand, we can probably follow a simple rule:
if it's not exported via bpf_flow_keys (and src/dsc mac is not),
tc_flower is not supported as well.