Hi Alexei,
I was syncing with Arnaldo offline on whether the change in Pahole was
enough for supporting percpu vars in BPF. While I think about it, I
felt it's better to have a flag in pahole to toggle the feature on and
off. My concern is mainly malformed BTF caused by the newly introduced
VAR entries. I've tested over some kernel configurations and compilers
but not confident enough.
I haven't thought about opt-out. Maybe we should use opt-out instead of opt-in?
Hao
On Fri, Sep 18, 2020 at 5:58 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Fri, Sep 18, 2020 at 1:41 PM Hao Luo <haoluo@xxxxxxxxxx> wrote:
> >
> > Previous commit f3d9054ba8ff ("btf_encoder: Teach pahole to store
> > percpu variables in vmlinux BTF.") introduced a feature in btf_encoder
> > that encodes global symbols in BTF. However, this feature is not
> > protected by any flag. In order to avoid surprises after Pahole v1.18
> > rolls out, make this feature off by default and enable only upon
> > request. On the kernel side, we can add a Kconfig to enable this
> > new capability on an opt-in pattern.
>
> I think I've missed some earlier discussions and motivation for opt-in.
> Why opt-out is a problem?
> I would think the kernel build should always include percpu vars.
