RE: [PATCH bpf-next v3 0/5] xdp: Support multiple programs on a single interface through chain calls

["Toke Høiland-Jørgensen" <toke@redhat.com>]

John Fastabend <john.fastabend@gmail.com> writes:

> Toke Høiland-Jørgensen wrote:
>> This series adds support for executing multiple XDP programs on a single
>> interface in sequence, through the use of chain calls, as discussed at the Linux
>> Plumbers Conference last month:
>> 
>> https://linuxplumbersconf.org/event/4/contributions/460/
>> 
>
> Can we add RFC to the title if we are just iterating through
> idea-space here.

I don't view this as "just iterating through idea-space". I'll admit
that I may have overestimated the extent to which we were all on the
same page on this after LPC, but I do view these submissions as serious
proposals that are making progress... :)

>> # HIGH-LEVEL IDEA
>> 
>> Since Alexei pointed out some issues with trying to rewrite the eBPF byte code,
>> let's try a third approach: We add the ability to chain call programs into the
>> eBPF execution core itself, but without rewriting the eBPF byte code.
>> 
>> As in the previous version, the bpf() syscall gets a couple of new commands
>> which takes a pair of BPF program fds and a return code. It will then attach the
>> second program to the first one in a structured keyed by return code. When a
>> program chain is thus established, the former program will tail call to the
>> latter at the end of its execution.
>> 
>> The actual tail calling is achieved by adding a new flag to struct bpf_prog and
>> having BPF_PROG_RUN run the chain call logic if that flag is set. This means
>> that if the feature is *not* used, the overhead is a single conditional branch
>> (which means I couldn't measure a performance difference, as can be seen in the
>> results below).
>> 
>
> I still believe user space should be able to link these multiple
> programs together as Ed and I were suggesting in the last series.

I expect that userspace probably could (I mean, after all, eBPF is
within spitting distance of a full almost-Turing-complete executing
environment so userspace can conceivably do pretty much anything).

However, I don't believe that doing it in userspace is the best
solution. I view it as a tradeoff: How much complexity do we have to add
to the kernel to make things easier for userspace. And given that we can
do this without negatively impacting performance, and with a reasonable
cost in terms of complexity (both of which I think this series
demonstrates that yes, we can), I think this makes sense to put in the
kernel.

> Also by doing it by linking your control program can be arbitrary
> complex. For example not just taking the output of one program and
> jumping to another but doing arbitrary more complex/interesting
> things. Taking the input from multiple programs to pick next call for
> example.

I expect there will indeed be more complex use cases where combining
multiple programs in arbitrary complex ways would make a lot of sense,
and doing that by linking in userspace is probably a good fit for that.
But for the simple use case of running multiple programs after one
another, I think it is overkill, and something that is better to do in
the kernel.

-Toke