Re: [RFC PATCH bpf-next 2/4] bpf: introduce BPF dispatcher

From: Edward Cree <ecree@solarflare.com>
To: "Björn Töpel" <bjorn.topel@gmail.com>
Cc: Netdev <netdev@vger.kernel.org>,
	"Alexei Starovoitov" <ast@kernel.org>,
	"Daniel Borkmann" <daniel@iogearbox.net>,
	"Björn Töpel" <bjorn.topel@intel.com>, bpf <bpf@vger.kernel.org>,
	"Magnus Karlsson" <magnus.karlsson@gmail.com>,
	"Karlsson, Magnus" <magnus.karlsson@intel.com>,
	"Jonathan Lemon" <jonathan.lemon@gmail.com>
Subject: Re: [RFC PATCH bpf-next 2/4] bpf: introduce BPF dispatcher
Date: Thu, 14 Nov 2019 10:18:55 +0000	[thread overview]
Message-ID: <96811723-ab08-b987-78c7-2c9f2a0a972c@solarflare.com> (raw)
In-Reply-To: <CAJ+HfNiDa912Uwt41_KMv+Z-sGr8fU7s4ncBPiUSx4PPAMQQqQ@mail.gmail.com>

On 14/11/2019 06:29, Björn Töpel wrote:
> On Wed, 13 Nov 2019 at 22:41, Edward Cree <ecree@solarflare.com> wrote:
>> On 13/11/2019 20:47, Björn Töpel wrote:
>> The first-come-first-served model for dispatcher slots might mean that
>>  a low-traffic user ends up getting priority while a higher-traffic
>>  user is stuck with the retpoline fallback.  Have you considered using
>>  a learning mechanism, like in my dynamic call RFC [1] earlier this
>>  year?  (Though I'm sure a better learning mechanism than the one I
>>  used there could be devised.)
> My rationale was that this mechanism would almost exclusively be used
> by physical HW NICs using XDP. My hunch was that the number of netdevs
> would be ~4, and typically less using XDP, so a more sophisticated
> mechanism didn't really make sense IMO.
That seems reasonable in most cases, although I can imagine systems with
 a couple of four-port boards being a thing.  I suppose the netdevs are
 likely to all have the same XDP prog, though, and if I'm reading your
 code right it seems they'd share a slot in that case.

> However, your approach is more
> generic and doesn't require any arch specific work. What was the push
> back for your work?
Mainly that I couldn't demonstrate a performance benefit from the few
 call sites I annotated, and others working in the area felt that
 manual annotation wouldn't scale — Nadav Amit had a different approach
 [2] that used a GCC plugin to apply a dispatcher on an opt-out basis
 to all the indirect calls in the kernel; the discussion on that got
 bogged down in interactions between text patching and perf tracing
 which all went *waaaay* over my head.  AFAICT the static_call series I
 was depending on never got merged, and I'm not sure if anyone's still
 working on it.

-Ed

[2] https://lkml.org/lkml/2018/12/31/19