[Buildroot] [PATCH v3 1/1] package/bpftool: revert bpf_cookie patch to allow building

[James Hilliard]

On Mon, Jun 20, 2022 at 12:45 AM Arnout Vandecappelle <arnout at mind.be> wrote:
>
>
>
> On 20/06/2022 01:19, James Hilliard wrote:
> > On Sun, Jun 19, 2022 at 9:20 AM Arnout Vandecappelle <arnout at mind.be> wrote:
> >>
> >>
> >>
> >> On 16/06/2022 10:11, Shahab Vahedi wrote:
> >>> On 6/16/22 01:27, James Hilliard wrote:
> >>>> On Wed, Jun 15, 2022 at 5:10 AM Shahab Vahedi via buildroot
> >>>> <buildroot at buildroot.org> wrote:
> >>>>>
> >>>>> On 6/14/22 19:14, Arnout Vandecappelle wrote:
> >>>>>>
> >>>>>> On 14/06/2022 11:31, Shahab Vahedi wrote:
> >>>>>>> Building bpftool on Debian 11 (bullseye) with kernel v5.10 and clang-11
> >>>>>>
> >>>>>>    How do you build host-bpftool with clang in Buildroot context? HOSTCC is set to gcc in the Makefile... Do you supply an explicit HOSTCC= on the Buildroot command line? I'm not sure if we are really interested in carrying fixes for such exotic and not-really-supported situations...
> >>>>>
> >>>>> No, I don't do any sort of trickery to build bpftool on my end. The
> >>>>> bootstrapping, if becomes available for your configuration, uses clang
> >>>>> and only clang. I tried to explain this in v4 of the patch [1], the
> >>>>> second paragraph of the commit message.
> >>>>
> >>>> I think clang/llvm support isn't going to work correctly yet since we only have
> >>>> version 9.0.1, there's a series bumping to version 11.1.0 that should fix that:
> >>>> https://patchwork.ozlabs.org/project/buildroot/list/?series=291585
> >>>>
> >>>> Minimum clang/llvm version for libbpf co-re is version 10:
> >>>> https://github.com/libbpf/libbpf*bpf-co-re-compile-once--run-everywhere
> >>>
> >>> You're right about the clang version, but it doesn't have anything to do
> >>> with Buildroot's clang. The build process uses the clang that is installed
> >>> on the host. For Debian bullseye, that is clang 11.
> >>   >
> >>> To emphasise, I am cross-building bpftool for my "arc-linux" target, and yet
> >>> the bootstrap part of bpftool, uses the x86 clang of the Debian machine.
> >>
> >>
> >>    Hm, that smells like we actually want to build host-clang (after updating it
> >> to 11.1.0 of course) so that we are sure a known version of clang is used.
> >> Possibly with a check-host-clang check to avoid building it if the installed
> >> clang is good enough.
> >
> > Dealing with multiple external clang versions may be a bit tricky and difficult
>
>   We do it for GCC, so we can do something similar for clang.
>
> > to test properly, we don't really want to use the system clang/llvm although
> > clang/llvm external toolchain support may be desirable here as those could
> > be tested by the autobuilders.
>
>   For GCC, the host toolchain is completely unrelated to the target toolchain.
> With clang, it's true that it's possible to use the target compiler for host
> builds as well, but don't we still need binutils? So in my mind, there would be
> a separate host clang toolchain and target clang toolchain.

BPF is kinda weird...for both clang and GCC. It's sorta a separate
architecture in
the compiler...but is also sorta not architecture specific in what it
builds for.

>
>
> > It would be good to get clang/llvm updated soon as systemd is now using bpf for
> > some service security/isolation features that we currently aren't able
> > to support
> > due to clang/llvm being too old.
>
>   Unfortunately your series is rather large and has no Reviewed or Acked by
> tags... So it tends to languish on patchwork.

The tested-by for the v12 series ended up in the wrong thread I think:
https://lore.kernel.org/buildroot/BN2P110MB16408DC4537E54EF7ECC7906F21A9@BN2P110MB1640.NAMP110.PROD.OUTLOOK.COM/

>
>
> >>    And we probably want a user-visible option to enable co-re then, because it's
> >> going to be expensive to build.
> >
> > We may want to make llvm/clang part of the pre-built toolchains
> > eventually, but for
> > now I'd say we should just conditionally enable co-re here if we are
> > already building
> > a clang/llvm toolchain.
>
>   Although I'm usually in favour of automatic dependencies, in this case I'd say
> it's worth adding an explicit config option.
>
>
> > There's some early GCC support for co-re in 12.1 which I was experimenting with
>
>   But then it would depend on both host and target GCC >= 12...

I thought we don't support GCC on the target.

We would essentially have a 2 architecture cross-toolchain(one real
target arch, and the
BPF virtual architecture compiler/assemblers and such).

>
> > as well which may reduce the need for a llvm/clang toolchain for co-re. The GCC
> > toolchain BPF support build process is a bit complex however as one needs to
> > sorta do a hybrid multitarget build with GCC and binutils since GCC
> > treats BPF as
> > a separate target(and since GCC along with the binutils GAS assembler
> > don't natively
> > handle multi-target toolchain builds themselves).
>
>   Ouch, so we still can't reuse the existing host toolchain for the host parts?
> then it doesn't help that much, does it?

We can create a GCC toolchain that can target both BPF and the normal target
architecture...but it's sorta a strange multiarch style toolchain.

>
>   Regards,
>   Arnout
>
> >
> > I have an experimental branch for that here:
> > https://github.com/buildroot/buildroot/compare/master...jameshilliard:ebpf
> >
> > I'll try and clean that up a bit once the GCC 12.1 series it is based
> > on is merged:
> > https://patchwork.ozlabs.org/project/buildroot/list/?series=302389
> >
> >>
> >>    Regards,
> >>    Arnout
> >>