[xdp-hints] Re: [PATCH bpf-next v3 11/12] mlx5: Support RX XDP metadata

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

On Thu, Dec 8, 2022 at 4:02 PM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
>
> Stanislav Fomichev <sdf@google.com> writes:
>
> > On Thu, Dec 8, 2022 at 2:59 PM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
> >>
> >> Stanislav Fomichev <sdf@google.com> writes:
> >>
> >> > From: Toke Høiland-Jørgensen <toke@redhat.com>
> >> >
> >> > Support RX hash and timestamp metadata kfuncs. We need to pass in the cqe
> >> > pointer to the mlx5e_skb_from* functions so it can be retrieved from the
> >> > XDP ctx to do this.
> >>
> >> So I finally managed to get enough ducks in row to actually benchmark
> >> this. With the caveat that I suddenly can't get the timestamp support to
> >> work (it was working in an earlier version, but now
> >> timestamp_supported() just returns false). I'm not sure if this is an
> >> issue with the enablement patch, or if I just haven't gotten the
> >> hardware configured properly. I'll investigate some more, but figured
> >> I'd post these results now:
> >>
> >> Baseline XDP_DROP:         25,678,262 pps / 38.94 ns/pkt
> >> XDP_DROP + read metadata:  23,924,109 pps / 41.80 ns/pkt
> >> Overhead:                   1,754,153 pps /  2.86 ns/pkt
> >>
> >> As per the above, this is with calling three kfuncs/pkt
> >> (metadata_supported(), rx_hash_supported() and rx_hash()). So that's
> >> ~0.95 ns per function call, which is a bit less, but not far off from
> >> the ~1.2 ns that I'm used to. The tests where I accidentally called the
> >> default kfuncs cut off ~1.3 ns for one less kfunc call, so it's
> >> definitely in that ballpark.
> >>
> >> I'm not doing anything with the data, just reading it into an on-stack
> >> buffer, so this is the smallest possible delta from just getting the
> >> data out of the driver. I did confirm that the call instructions are
> >> still in the BPF program bytecode when it's dumped back out from the
> >> kernel.
> >>
> >> -Toke
> >>
> >
> > Oh, that's great, thanks for running the numbers! Will definitely
> > reference them in v4!
> > Presumably, we should be able to at least unroll most of the
> > _supported callbacks if we want, they should be relatively easy; but
> > the numbers look fine as is?
>
> Well, this is for one (and a half) piece of metadata. If we extrapolate
> it adds up quickly. Say we add csum and vlan tags, say, and maybe
> another callback to get the type of hash (l3/l4). Those would probably
> be relevant for most packets in a fairly common setup. Extrapolating
> from the ~1 ns/call figure, that's 8 ns/pkt, which is 20% of the
> baseline of 39 ns.
>
> So in that sense I still think unrolling makes sense. At least for the
> _supported() calls, as eating a whole function call just for that is
> probably a bit much (which I think was also Jakub's point in a sibling
> thread somewhere).

imo the overhead is tiny enough that we can wait until
generic 'kfunc inlining' infra is ready.

We're planning to dual-compile some_kernel_file.c
into native arch and into bpf arch.
Then the verifier will automatically inline bpf asm
of corresponding kfunc.