Re: [PATCH bpf-next RFC v1 3/8] libbpf: Support kernel module function calls

On Tue, Aug 31, 2021 at 7:59 PM Alexei Starovoitov
[off-list ref] wrote:

On Tue, Aug 31, 2021 at 7:27 PM Kumar Kartikeya Dwivedi
[off-list ref] wrote:

quoted

quoted

quoted

@@ -5327,6 +5340,7 @@ bpf_object__relocate_data(struct bpf_object *obj, struct bpf_program *prog)
                        ext = &obj->externs[relo->sym_off];
                        insn[0].src_reg = BPF_PSEUDO_KFUNC_CALL;
                        insn[0].imm = ext->ksym.kernel_btf_id;
+                       insn[0].off = ext->ksym.offset;

Just a few lines above we use insn[1].imm =
ext->ksym.kernel_btf_obj_fd; for EXT_KSYM (for variables). Why are you
inventing a new form if we already have a pretty consistent pattern?

That makes sense. This is all new to me, so I went with what was described in
e6ac2450d6de (bpf: Support bpf program calling kernel function), but I'll rework
it to encode the btf fd like that in the next spin. It also makes the everything
far simpler.

Hmm. kfunc call is a call insn. There is no imm[1].

Doh, right :( Never mind, we'll need to use fd_array for this.

Either way, I don't think hashmap use is warranted here to find a BTF
slot. Let's just do linear search, it's not like we are going to have
thousands of module BTFs used by any single BPF program, right?