On Sat, Jul 27, 2019 at 10:00 AM Alexei Starovoitov <ast@xxxxxx> wrote:
>
> On 7/26/19 11:25 PM, Andrii Nakryiko wrote:
> >>> + } else if (class == BPF_ST && BPF_MODE(insn->code) == BPF_MEM) {
> >>> + if (insn->imm != orig_off)
> >>> + return -EINVAL;
> >>> + insn->imm = new_off;
> >>> + pr_debug("prog '%s': patched insn #%d (ST | MEM) imm %d -> %d\n",
> >>> + bpf_program__title(prog, false),
> >>> + insn_idx, orig_off, new_off);
> >> I'm pretty sure llvm was not capable of emitting BPF_ST insn.
> >> When did that change?
> > I just looked at possible instructions that could have 32-bit
> > immediate value. This is `*(rX) = offsetof(struct s, field)`, which I
> > though is conceivable. Do you think I should drop it?
>
> Just trying to point out that since it's not emitted by llvm
> this code is likely untested ?
> Or you've created a bpf asm test for this?
Yeah, it's untested right now. Let me try to come up with LLVM
assembly + relocation (not yet sure how/whether builtin works with
inline assembly), if that works out, I'll leave this, if not, I'll
drop BPF_ST|BPF_MEM part.
>
>
