Re: [RFC bpf-next v3 6/8] flow_dissector: handle no-skb use case

[Stanislav Fomichev <sdf@fomichev.me>]

On 03/28, Willem de Bruijn wrote:
> > > > > > > > > > If skb_vlan_tag_present(skb) returns true, we set proto to skb->protocol
> > > > > > > > > > and move on.
> > > > > > > > > >
> > > > > > > > > > But, we would need vlan_proto/present/tci in the flow_keys in the future.
> > > > > > > > > > We don't currently return parsed vlan data from the BPF flow dissector.
> > > > > > > > > > But it feels like it's getting into bpf-next territory :-)
> > > > > > > > >
> > > > > > > > > Whether ctx->data points to L2 or L3 is uapi regardless whether
> > > > > > > > > progs/bpf_flow.c is relying on that or not.
> > > > > > > > > So far I think you're saying that in all three cases:
> > > > > > > > > no-skb, skb befor rfs, skb after rfs ctx->data points to L2, right?
> > > > > > > > > This has to be preserved.
> > > > > > > > It points to L3 (or vlan). And this will be preserved, I have no
> > > > > > > > intention to change that.
> > > > > > > >
> > > > > > > > Just to make sure, we are on the same page, here is what
> > > > > > > > __skb_flow_dissect (and BPF prog) is seeing in nhoff.
> > > > > > > >
> > > > > > > > NO-VLAN is always the same for both with-skb/no-skb:
> > > > > > > > +----+----+-----+--+
> > > > > > > > |DMAC|SMAC|PROTO|L3|
> > > > > > > > +----+----+-----+--+
> > > > > > > >                  ^
> > > > > > > >                  +-- nhoff
> > > > > > > >                      proto = PROTO
> > > > > > > >
> > > > > > > > VLAN no-skb (eth_get_headlen):
> > > > > > > > +----+----+----+---+-----+--+
> > > > > > > > |DMAC|SMAC|TPID|TCI|PROTO|L3|
> > > > > > > > +----+----+----+---+-----+--+
> > > > > > > >                 ^
> > > > > > > >                 +-- nhoff
> > > > > > > >                     proto = TPID
> > > > > > >
> > > > > > > where ctx->data will point to ?
> > > > > > > These nhoff differences are fine.
> > > > > > > I want to make sure that ctx->data is the same for all.
> > > > > > For with-skb, nhoff would be zero, and ctx->data would point to
> > > > > > TCI/L3.
> > > > > > For skb-less, ctx->data would point to L2 (DMAC), and nhoff would be
> > > > > > non-zero (TCI/L3 offset).
> > > > > >
> > > > > > If you want, for skb-less case, when calling BPF program we can do the math
> > > > > > ourselves and set ctx->data to data + nhoff, and pass nhoff = 0.
> > > > > > But I'm not sure whether we need to do that; flow dissector is supposed
> > > > > > to look at ctx->data + nhoff, it should not matter what each individual
> > > > > > value is, they only make sense together.
> > > > >
> > > > > My strong preference is to have data to point to L2 in all cases.
> > > > > Semantics of requiring bpf prog to start processing from a tuple
> > > > > (data + nhoff) where both point to random places is very confusing.
> > > >
> > > > Since flow dissection starts at the network layer, I would then
> > > > suggest data always at L3 and nhoff 0.
> > For eth_get_headlen we need to manually parse 802.1q header. And for RFS
> > case as well (unless I'm missing something).
> >
> > > > This can be derived in the same manner as __skb_flow_dissect
> > > > already does if !data, using only skb_network_offset.
> > > >
> > > > From a quick scan, skb_mac_offset should also be valid in all cases
> > > > where the flow dissector is called today, so the other can be computed, too.
> > > >
> > > > But this is less obvious. For instance, tun_get_user calls into the flow
> > > > dissector up to three times (wow) and IFF_TUN has no link layer
> > > > (ARPHRD_NONE). And then there are also fun variable length link layer
> > > > protocols to deal with..
> > >
> > > ahh. ok. Can we guarantee some stable position?
> > I don't think so. Pre RFS ctx->data+nhoff can point to 802.1q header,
> > post RFS it will point to L3. The only thing we can do is to have
> > nhoff=0 (and adjust ctx->data accordingly) when the main bpf
> > flow dissector procedure is called. But that would require bringing
> > this new kernel context (bpf_flow_dissector) into bpf/stable.
> > (And it's not clear what's the benefit, since tail calls would still
> > have to look at that offset).
> 
> The flow dissector can be called also before and after tunneling, in
> which case skb_network_offset points to an inner header. Or after
> MPLS, which stumps a flow dissector called earlier as that has no
> information about the encapsulated protocol.
> 
> I don't think that there should be a goal that flow dissection starts
> at the same point in the packet for all callsites along the datapath.
> As long as it always starts at a known ETH_P_.. type protocol header
> the program should be able to parse that. That is how the non-BPF
> flow dissector works.
> 
> > > Current bpf_flow_dissect_get_header assumes that
> > > ctx->data + ctx->flow_keys->thoff point to IP, right?
> > Yes, mostly, except that if skb->protocol is 802.1q/ad, it's 802.1q header.
> > And it's only for the "main" call; bpf program adjusts this thoff
> > to make sure that tail calls preserve some sense of progress (so it
> > eventually points to L4 and that's what we export back).
> >
> > > Based on what Stanislav saying above even that is not a guarantee?
> > > I'm struggling to see how users can wrap their heads around this.
> > > It seems bpf_flow.c will become the only prog that can deal with
> > > this range of possible inputs.
> > >
> > > I propose to start with the doc that describes all cases, where
> > > things point to and how prog suppose to parse that.
> > Yeah, that is what I was going to propose - add a doc along with the
> > patch series. I don't see how we can make it simple(r) at this point :-(
> 
> Does it have to be simpler? A flow dissector should be ready to
> dissect VLAN tags. That's the only complication here?
I don't see how it can be made simpler. That's the context from which
existing __skb_flow_dissect is called and that's what we have to dissect
from the BPF as well. We can try to make nhoff to be 0 when the
dissector is called, that's probably the only simplification we can
attempt to do (but, as I said previously, it requires bringing
new kernel context to bpf/stable and seems more complicated than
necessary).

Let me prepare a series for bpf/stable with the small doc describing
BPF flow dissector environment. We can continue the discussion
from there :-)

> > I can try to document everything so users don't have to read the
> > kernel code to understand how to write the bpf flow dissector programs.