Re: [PATCH RFC PoC 0/3] nftables meets bpf

Hi Daniel,

On Mon, Feb 19, 2018 at 08:57:39PM +0100, Daniel Borkmann wrote:

On 02/19/2018 05:37 PM, Pablo Neira Ayuso wrote:
[...]

quoted

* Simplified infrastructure: We don't need the ebpf verifier complexity
  either given we trust the code we generate from the kernel. We don't
  need any complex userspace tooling either, just libnftnl and nft
  userspace binaries.

* Hardware offload: We can use this to offload rulesets to the only
  smartnic driver that we have in the tree that already implements bpf
  offload, hence, we can reuse this work already in place.

In addition Dave's points, regarding the above two, this will also only
work behind the verifier since NIC offloading piggy-backs on the verifier's
program analysis to prepare and generate a dev specific JITed BPF
prog, so it's not the same as normal host JITs (and there, the cBPF ->
eBPF in kernel migration adds a lot of headaches already due to
different underlying assumptions coming from the two flavors, even
if both are eBPF insns in the end), and given this, offloading will
also only work for eBPF and not cBPF.

We also have a large range of TCAM based hardware offload outthere
that will _not_ work with your BPF HW offload infrastructure. What
this bpf infrastructure pushes into the kernel is just a blob
expressing things in a very low-level instruction-set: trying to find
a mapping of that to typical HW intermediate representations in the
TCAM based HW offload world will be simply crazy.

There's a lot more the verifier is doing internally, like performing
various different program rewrites from the context, for helpers
(e.g. inlining), and for internal insn mappings that are not exposed
(e.g. in calls), so we definitely need to go through it.

If we need to call the verifier from the kernel for the code that we
generate there for this initial stage, that should be not an issue.

The BPF interface is lacking many of the features and flexibility we
have in netlink these days, and it is only allowing for monolitic
ruleset replacement. This approach also loses internal rule stateful
information that we're doing in the packet path when updating the
ruleset. So it's taking us back to exactly the same mistakes we made
in iptables back in the 90s as it's been mentioned already.

So I just wish I can count with your help in this process, we can get
the best of the two worlds by providing a subsystem that allows users
to configure packet classification through one single interface, no
matter if the policy representation ends up being in software or HW
offloads, either TCAM or smartnic.

Thanks.