Re: libbpf missing set_link_xdp_fd ?

[Daniel Borkmann <daniel@iogearbox.net>]

On 09/25/2017 06:01 PM, Alexei Starovoitov wrote:
> On Mon, Sep 25, 2017 at 01:02:04PM +0200, Daniel Borkmann wrote:
>> On 09/25/2017 11:41 AM, Daniel Borkmann wrote:
>>> On 09/25/2017 11:25 AM, Jesper Dangaard Brouer wrote:
>>>> On Sun, 24 Sep 2017 20:07:11 +0200
>>>> Eric Leblond <eric@regit.org> wrote:
>>>>
>>>>> Hello,
>>>>>
>>>>> bpf_load is implementing the set_link_xdp_fd() function and it seems
>>>>> this function is not available in libbpf.
>>>>
>>>> I also ran into this problem, when I wanted to switch my XDP programs
>>>> into using tools/lib/bpf/.
>>>
>>> Right.
>>>
>>>> E.g.: https://github.com/netoptimizer/prototype-kernel/commit/5397e596c30416763
>>>>
>>>>> Should it be the case or should it be part of another library ?
>>>>
>>>> I'm not sure... maybe Alexei, Daniel or Wang have an opinion?
>>>
>>> It seems that none of the actual attach infra is handled in
>>> tools/lib/bpf/, meaning neither attaching the fd to kprobes/
>>> uprobes/tracepoints nor tc or xdp, so adding a special case
>>> only for xdp into the lib feels that we probably might be better
>>> off putting this either into its own mini lib on top of that,
>>
>> That could potentially go to tools/lib/xdp/, probably would be
>> the best option.
>
> I'm afraid it may spiral out of control.
> If we create such dirs, we'd need one for xdp, tc, cgroup, tracepoint, kprobe.
> Differentiating the libraries by type of attach point is imo too fine grained.
>
> Why not to put everything into toos/lib/bpf ?
> New file for netlink stuff is easy enough and we can gatekeep it with
> some knob in the Makefile, so perf build won't need to include it.

Personally, I don't really have a strong opinion on that. This could
probably be made a bit modular for each attach handler with the option
to compile some of them out if not desired, so could work as well.