Re: [PATCH bpf-next 2/2] libbpf: Implement API for generating BTF for ebpf objects

From: Andrii Nakryiko <andrii.nakryiko@gmail.com>
To: "Mauricio Vásquez Bernal" <mauricio@kinvolk.io>
Cc: Networking <netdev@vger.kernel.org>, bpf <bpf@vger.kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Andrii Nakryiko <andrii@kernel.org>,
	Rafael David Tinoco <rafaeldtinoco@gmail.com>,
	Rafael David Tinoco <rafael.tinoco@aquasec.com>,
	Lorenzo Fontana <lorenzo.fontana@elastic.co>
Subject: Re: [PATCH bpf-next 2/2] libbpf: Implement API for generating BTF for ebpf objects
Date: Thu, 28 Oct 2021 15:48:05 -0700	[thread overview]
Message-ID: <CAEf4BzZ5QmHwD-FNgqBsJWb2j+7iZfTNe_+E6vAS-HUrwUDmjw@mail.gmail.com> (raw)
In-Reply-To: <CAHap4zt1jFM_hMd0mqT+158f3-C8Vn0AtZHH+pK_MxxiUan5zg@mail.gmail.com>

On Thu, Oct 28, 2021 at 3:42 PM Mauricio Vásquez Bernal
<mauricio@kinvolk.io> wrote:
>
> > I don't think it's necessary for libbpf to expose all these new APIs.
> > The format of CO-RE relocations and .BTF.ext is open and fixed. You
> > don't really need to simulate full CO-RE relocation logic to figure
> > out which types are necessary. Just go over all .BTF.ext records for
> > CO-RE relocations, parse spec (simple format as well) and see which
> > fields are accessed.
>
> How do you suggest to match the types for the target BTF without
> simulating the CO-RE relocation? Are you suggesting to match them only
> by name? We want to generate the minimal BTF that is needed by a given
> object. Consider that we could generate these files for thousands of
> kernels, size is very important for us. For this reason we chose to
> simulate the relocation generating only the types (and members) that
> are really needed.
>

How many unnecessary structs are matching if you match by name only?

Keep in mind, if your kernel BTF has task_struct and task_struct___2,
then CO-RE relocation will keep matching both; and that's not an error
for libbpf if all the field offsets will be consistent.

In short, I think simple name matching for trimming down BTF is
completely adequate. CO-RE relocation has to be more strict about
matching, but the subset of types that are used will be the same or
almost the same.

> > Either way, this is not libbpf's problem to solve. It's a tooling problem.
>
> I agree. When I started working on this I tried to implement it
> without using the libbpf relocation logic, but very soon I realized I
> was  reimplementing the same logic. Another possibility we have
> considered is to expose this relocation logic in the libbpf API,
> however I fear it's too complicated and invasive too...