Re: pahole + BTF was: Re: [Question] bpf: about a new 'tools/bpf/bpf_dwarf2btf'

[Daniel Borkmann]

On 07/25/2018 10:11 PM, Martin KaFai Lau wrote:
> On Thu, Jul 26, 2018 at 04:21:31AM +0900, Taeung Song wrote:
>> On 07/26/2018 03:27 AM, Taeung Song wrote:
>>> On 07/26/2018 02:52 AM, Arnaldo Carvalho de Melo wrote:
>>>> Em Thu, Jul 26, 2018 at 02:23:32AM +0900, Taeung Song escreveu:
>>>>> Hi,
>>>>>
>>>>> Building bpf programs with .BTF section,
>>>>> I thought it'd be better to convert dwarf info to .BTF by
>>>>> a new tool such as 'tools/bpf/bpf_dwarf2btf' instead of pahole
>>>>> in the future.
>>>>> Currently for bpf binary that have .BTF section,
>>>>> we need to use pahole from https://github.com/iamkafai/pahole/tree/btf
>>>>> with the command line such as "pahole -J bpf_prog.o".
>>>>> I think it is great but if implementing new 'bpf_dwarf2btf'
>>>>> (dwarf parsing + btf encoder code written by Martin KaFai Lau on
>>>>> the pahole project i.e. btf.h, btf_encoder.c, btf_encoder.h,
>>>>> libbtf.c, libbtf.h),
>>>>> BPF developers would more easily use functionalities based on BTF.
>>>>
>>>> What would be easier exactly? Not having to install a package but build
>>>> it from the kernel sources?
>>>>
>>>> Many kernel developers already have pahole installed for other uses, so
>>>> no need to install anything.
>>>
>>> Understood, but I think there are many non-kernel developers
>>> developing BPF programs and they mightn't have or use pahole.
>>>
>>> So, if providing the 'dwarf2btf' feature on tools/bpf or tools/bpf/bpftool,
>>> non-kernel developers can also more easily build bpf prog with .BPF, no ?
> Some quick thoughts,
> IMO, I suspect if it is in the distro's pahole package,  it should be easy
> enough for kernel and non kernel developer to install.
> BTF usage is still evolving,  we might re-evaluate going forward but at this
> point I think leveraging pahole's existing capability is a good option.

Agree, if there will be a future use-case where pahole might not be well-fitting,
we could add it to bpftool then so I wouldn't rule it out, but for the functionality
right now it seems good to reuse it. Presumably BPF developers have it installed
anyway to inspect struct padding from BPF obj files.

Thanks,
Daniel