Re: [PATCH RFC bpf-next 0/7] bpf: Support struct_ops in light skeletons

[Anton Protopopov]

On Mon, Nov 9, 2026 at 10:40 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Mon, Nov 09, 2026 at 11:46 AM Ihor Solodrai <ihor.solodrai@xxxxxxxxx> wrote:
>
> > Andrii, Alexei, wdyt? Any other ideas?
>
> The loader asks the kernel for ids, sizes and offsets only to pass them
> back in the next three commands:
>
>   the loader asks                   and passes it back in
>   --------------------------------  ------------------------------------
>   id, size of bpf_struct_ops_<t>    BPF_MAP_CREATE
>   id of <t>, index of the member    BPF_PROG_LOAD
>   offset of data, of every member   BPF_MAP_UPDATE_ELEM
>
> the kernel can match by name instead.
>
> The kernel has both BTFs. BPF_MAP_CREATE of a struct_ops map demands
> btf_fd of the object already, and bpf_struct_ops.c doesn't use it.
> The kernel can find its struct_ops and the members by name,
> like btf_parse_kptr() finds the type of a kptr.
> Then the loader passes what bpftool knows when it generates
> the skeleton:
>
>   command              the loader passes
>   -------------------  ---------------------------------------------------
>   BPF_MAP_CREATE       btf_fd, btf_value_type_id and value_size of the
>                        struct in the BTF of the object
>   BPF_PROG_LOAD        attach_btf_obj_fd = btf_fd, attach_btf_id = that
>                        struct, expected_attach_type = index of its member
>   BPF_MAP_UPDATE_ELEM  that struct with the prog fds in it
>
> The kernel rejects such BPF_MAP_CREATE and BPF_PROG_LOAD today.
> Nothing changes for existing users.
> The BTF of the object is in the blob, so the signature covers the names.
>
> Then no need for a helper or a kfunc. The 4k buffer, the fds of module
> BTF and the offsets at run time go away.
> map->def.value_size, prog->attach_btf_id and prog->expected_attach_type
> don't hold offsets into the blob.
>
> The diff below replaces patches 1-5. Patches 6 and 7 stay.

I like Alexei's approach here, it fits with shirudo. I also have tested it.
(Patches 6,7 weren't tested, shirudo uses libbpf without bpftool.)

>
> [...]
>