Re: XDP-hints: Howto support multiple BTF types per packet basis?

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Thu, Jun 24, 2021 at 6:08 AM Magnus Karlsson
<magnus.karlsson@gmail.com> wrote:
> >
> > and libbpf could do relocations based on the different meta structs,
> > even removing the code for the ones that don't exist on the running
> > kernel.
>
> Just wondering how this will carry over to user-space and AF_XDP since
> it sees the same metadata area as XDP? AFAIK, dynamic linkers today
> cannot relocate structs or remove members, but I am not up-to-date
> with the latest here so might be completely wrong. And it would be
> good not to have to recompile a user-space binary just because a new
> NIC came out with a new BTF ID and layout, but with the same metadata
> member name and format as previous NICs/BTF IDs. But I do not know how
> to solve these things in user-space at the moment (except to have
> fixed locations for a common set of metadata, but that is what we are
> trying to avoid), so any hints and suggestions are highly appreciated.

CO-RE is not a kernel only feature.
The BTF tests in selftest/bpf exercise most of CO-RE purely in user space.
The libbpf needs to know the original BTF and the target BTF to
match one to another.
One option for AF_XDP would be to write a bpf program to be run in user space
and let libbpf handle relocations.
Another option is to teach llvm x86 backend to support
__attribute__((preserve_access_index)).
The work done by llvm BPF backend can be copy pasted to x86 backend.
Then standard x86 binaries will support dynamic struct layout.
imo CO-RE for x86 backend would be great to do regardless of xdp hints.
It's a straightforward copy-paste. Only need to convince x86 llvm maintainers.