Re: RFC: Fixing SK_REUSEPORT from sk_lookup_* helpers

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Tue, May 21, 2019 at 04:47:58PM +0100, Lorenz Bauer wrote:
> On Fri, 17 May 2019 at 00:38, Nitin Hande <nitin.hande@gmail.com> wrote:
> >
> > On Thu, May 16, 2019 at 2:57 PM Alexei Starovoitov
> > <alexei.starovoitov@gmail.com> wrote:
> > >
> > > On Thu, May 16, 2019 at 09:41:34AM +0100, Lorenz Bauer wrote:
> > > > On Wed, 15 May 2019 at 18:16, Joe Stringer <joe@isovalent.com> wrote:
> > > > >
> > > > > On Wed, May 15, 2019 at 8:11 AM Lorenz Bauer <lmb@cloudflare.com> wrote:
> > > > > >
> > > > > > In the BPF-based TPROXY session with Joe Stringer [1], I mentioned
> > > > > > that the sk_lookup_* helpers currently return inconsistent results if
> > > > > > SK_REUSEPORT programs are in play.
> > > > > >
> > > > > > SK_REUSEPORT programs are a hook point in inet_lookup. They get access
> > > > > > to the full packet
> > > > > > that triggered the look up. To support this, inet_lookup gained a new
> > > > > > skb argument to provide such context. If skb is NULL, the SK_REUSEPORT
> > > > > > program is skipped and instead the socket is selected by its hash.
> > > > > >
> > > > > > The first problem is that not all callers to inet_lookup from BPF have
> > > > > > an skb, e.g. XDP. This means that a look up from XDP gives an
> > > > > > incorrect result. For now that is not a huge problem. However, once we
> > > > > > get sk_assign as proposed by Joe, we can end up circumventing
> > > > > > SK_REUSEPORT.
> > > > >
> > > > > To clarify a bit, the reason this is a problem is that a
> > > > > straightforward implementation may just consider passing the skb
> > > > > context into the sk_lookup_*() and through to the inet_lookup() so
> > > > > that it would run the SK_REUSEPORT BPF program for socket selection on
> > > > > the skb when the packet-path BPF program performs the socket lookup.
> > > > > However, as this paragraph describes, the skb context is not always
> > > > > available.
> > > > >
> > > > > > At the conference, someone suggested using a similar approach to the
> > > > > > work done on the flow dissector by Stanislav: create a dedicated
> > > > > > context sk_reuseport which can either take an skb or a plain pointer.
> > > > > > Patch up load_bytes to deal with both. Pass the context to
> > > > > > inet_lookup.
> > > > > >
> > > > > > This is when we hit the second problem: using the skb or XDP context
> > > > > > directly is incorrect, because it assumes that the relevant protocol
> > > > > > headers are at the start of the buffer. In our use case, the correct
> > > > > > headers are at an offset since we're inspecting encapsulated packets.
> > > > > >
> > > > > > The best solution I've come up with is to steal 17 bits from the flags
> > > > > > argument to sk_lookup_*, 1 bit for BPF_F_HEADERS_AT_OFFSET, 16bit for
> > > > > > the offset itself.
> > > > >
> > > > > FYI there's also the upper 32 bits of the netns_id parameter, another
> > > > > option would be to steal 16 bits from there.
> > > >
> > > > Or len, which is only 16 bits realistically. The offset doesn't really fit into
> > > > either of them very well, using flags seemed the cleanest to me.
> > > > Is there some best practice around this?
> > > >
> > > > >
> > > > > > Thoughts?
> > > > >
> > > > > Internally with skbs, we use `skb_pull()` to manage header offsets,
> > > > > could we do something similar with `bpf_xdp_adjust_head()` prior to
> > > > > the call to `bpf_sk_lookup_*()`?
> > > >
> > > > That would only work if it retained the contents of the skipped
> > > > buffer, and if there
> > > > was a way to undo the adjustment later. We're doing the sk_lookup to
> > > > decide whether to
> > > > accept or forward the packet, so at the point of the call we might still need
> > > > that data. Is that feasible with skb / XDP ctx?
> > >
> > > While discussing the solution for reuseport I propose to use
> > > progs/test_select_reuseport_kern.c as an example of realistic program.
> > > It reads tcp/udp header directly via ctx->data or via bpf_skb_load_bytes()
> > > including payload after the header.
> > > It also uses bpf_skb_load_bytes_relative() to fetch IP.
> > > I think if we're fixing the sk_lookup from XDP the above program
> > > would need to work.
> > >
> > > And I think we can make it work by adding new requirement that
> > > 'struct bpf_sock_tuple *' argument to bpf_sk_lookup_* must be
> > > a pointer to the packet and not a pointer to bpf program stack.
> > > Then helper can construct a fake skb and assign
> > > fake_skb->data = &bpf_sock_tuple_arg.sport
> > > It can check that struct bpf_sock_tuple * pointer is within 100-ish bytes
> > > from xdp->data and within xdp->data_end
> > > This way the reuseport program's assumption that ctx->data points to tcp/udp
> > > will be preserved and it can access it all including payload.
> > >
> > > This approach doesn't need to mess with xdp_adjust_head and adjust uapi to pass length.
> > > Existing progs/test_sk_lookup_kern.c will magically start working with XDP
> > > even when reuseport prog is attached.
> > > Thoughts?
> >
> > I like this approach. A fake_skb approach will normalize the bpf_sk_lookup_*()
> > API peering into the kernel API between TC and XDP invocation. Just one question
> > that comes, I remember one of the comments I received during my XDP commit
> > was the stateless nature of XDP services and providing a fake_skb may bring
> > some potential side-effects to the desire of statelessness. Is that
> > still a possibility?
> > How do we guard against it?
> 
> To follow up on this, I'm also not sure how to tackle a "fake skb". If
> I remember this
> came up during the flow dissector series, and wasn't met with
> enthusiasm. Granted,
> replacing the skb argument to the lookup functions seems even harder, so maybe
> this is the lesser evil?

flow_dissector pretends to have 'skb' as bpf prog argument.
It doesn't allocate actual 'struct sk_buff' on the kernel side.
The verifier rewrites __sk_buff->foo accesses into
'struct bpf_flow_dissector'->bar access.
I was hoping similar idea can apply here.
BPF_PROG_TYPE_SK_REUSEPORT type takes 'struct sk_reuseport_md *'
so we don't need real skb there anyway.
But for older socket_filter based so_reuseport progs it's more difficult,
since at the verification time they're generic socket_filter progs.
Right now I don't have a better idea than to do static per_cpu struct sk_buff.