Re: [PATCH 0/3] perf/bpftool: Allow to link libbpf dynamically

[Daniel Borkmann <daniel@iogearbox.net>]

On 11/27/19 9:24 PM, Andrii Nakryiko wrote:
> On Wed, Nov 27, 2019 at 8:38 AM Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
>> On Wed, Nov 27, 2019 at 1:48 AM Jiri Olsa <jolsa@kernel.org> wrote:
>>>
>>> hi,
>>> adding support to link bpftool with libbpf dynamically,
>>> and config change for perf.
>>>
>>> It's now possible to use:
>>>    $ make -C tools/bpf/bpftool/ LIBBPF_DYNAMIC=1
>>>
>>> which will detect libbpf devel package with needed version,
>>> and if found, link it with bpftool.
>>>
>>> It's possible to use arbitrary installed libbpf:
>>>    $ make -C tools/bpf/bpftool/ LIBBPF_DYNAMIC=1 LIBBPF_DIR=/tmp/libbpf/
>>>
>>> I based this change on top of Arnaldo's perf/core, because
>>> it contains libbpf feature detection code as dependency.
>>> It's now also synced with latest bpf-next, so Toke's change
>>> applies correctly.
>>
>> I don't like it.
>> Especially Toke's patch to expose netlink as public and stable libbpf api.
>> bpftools needs to stay tightly coupled with libbpf (and statically
>> linked for that reason).
>> Otherwise libbpf will grow a ton of public api that would have to be stable
>> and will quickly become a burden.

+1, and would also be out of scope from a BPF library point of view.

> I second that. I'm currently working on adding few more APIs that I'd
> like to keep unstable for a while, until we have enough real-world
> usage (and feedback) accumulated, before we stabilize them. With
> LIBBPF_API and a promise of stable API, we are going to over-stress
> and over-design APIs, potentially making them either too generic and
> bloated, or too limited (and thus become deprecated almost at
> inception time). I'd like to take that pressure off for a super-new
> and in flux APIs and not hamper the progress.
> 
> I'm thinking of splitting off those non-stable, sort-of-internal APIs
> into separate libbpf-experimental.h (or whatever name makes sense),
> and let those be used only by tools like bpftool, which are only ever
> statically link against libbpf and are ok with occasional changes to
> those APIs (which we'll obviously fix in bpftool as well). Pahole
> seems like another candidate that fits this bill and we might expose
> some stuff early on to it, if it provides tangible benefits (e.g., BTF
> dedup speeds ups, etc).
> 
> Then as APIs mature, we might decide to move them into libbpf.h with
> LIBBPF_API slapped onto them. Any objections?

I don't think adding yet another libbpf_experimental.h makes sense, it feels
too much of an invitation to add all sort of random stuff in there. We already
do have libbpf.h and libbpf_internal.h, so everything that does not relate to
the /stable and public/ API should be moved from libbpf.h into libbpf_internal.h
such as the netlink helpers, as one example, and bpftool can use these since
in-tree changes also cover the latter just fine. So overall, same page, just
reuse/improve libbpf_internal.h instead of a new libbpf_experimental.h.

Thanks,
Daniel