From: Stanislav Fomichev <sdf@xxxxxxxxxx>
Date: Mon, May 13, 2019 at 7:21 PM
To: Stanislav Fomichev
Cc: Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx>, Network
Development, bpf, David Miller, Alexei Starovoitov, Daniel Borkmann,
Willem de Bruijn <willemb@xxxxxxxxxx>, Petar Penkov
> > On 05/13, Willem de Bruijn wrote:
> > > On Mon, May 13, 2019 at 5:21 PM Willem de Bruijn
> > > <willemdebruijn.kernel@xxxxxxxxx> wrote:
> > > >
> > > > On Mon, May 13, 2019 at 5:02 PM Stanislav Fomichev <sdf@xxxxxxxxxxx> wrote:
> > > > >
> > > > > On 05/13, Willem de Bruijn wrote:
> > > > > > On Mon, May 13, 2019 at 3:53 PM Stanislav Fomichev <sdf@xxxxxxxxxx> 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.
But it is not guaranteed to be an Ethernet link layer device. Making
this a good example of why when moving to BPF for safety we should not
keep any C dissection code in the path at all.
> 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.
Agreed. I was using that as point of reference just now, too.
