Re: [PATCH RFC bpf-next v1 0/6] Introduce unstable CT lookup helpers

[Kumar Kartikeya Dwivedi <memxor@gmail.com>]

On Wed, Nov 03, 2021 at 04:46:42AM IST, Alexei Starovoitov wrote:
> On Sat, Oct 30, 2021 at 08:16:03PM +0530, Kumar Kartikeya Dwivedi wrote:
> > This series adds unstable conntrack lookup helpers using BPF kfunc support.  The
> > patch adding the lookup helper is based off of Maxim's recent patch to aid in
> > rebasing their series on top of this, all adjusted to work with kfunc support
> > [0].
> >
> > This is an RFC series, as I'm unsure whether the reference tracking for
> > PTR_TO_BTF_ID will be accepted.
>
> Yes. The patches look good overall.
> Please don't do __BPF_RET_TYPE_MAX signalling. It's an ambiguous name.
> _MAX is typically used for a different purpose. Just give it an explicit name.
> I don't fully understand why that skip is needed though.

I needed a sentinel to skip return type checking (otherwise check that return
type and prototype match) since existing kfunc don't have a
get_kfunc_return_type callback, but if we add bpf_func_proto support to kfunc
then we can probably convert existing kfuncs to that as well and skip all this
logic. Mostly needed it for RET_PTR_TO_BTF_ID_OR_NULL.

Extending to support bpf_func_proto seemed like a bit of work so I wanted to get
some feedback first on all this, before working on it.

> Why it's not one of existing RET_*. Duplication of return and
> being lazy to propagate the correct ret value into get_kfunc_return_type ?
>
> > If not, we can go back to doing it the typical
> > way with PTR_TO_NF_CONN type, guarded with #if IS_ENABLED(CONFIG_NF_CONNTRACK).
>
> Please don't. We already have a ton of special and custom types in the verifier.
> refcnted PTR_TO_BTF_ID sounds as good way to scale it.
>

Understood.

> > Also, I want to understand whether it would make sense to introduce
> > check_helper_call style bpf_func_proto based argument checking for kfuncs, or
> > continue with how it is right now, since it doesn't seem correct that PTR_TO_MEM
> > can be passed where PTR_TO_BTF_ID may be expected. Only PTR_TO_CTX is enforced.
>
> Do we really allow to pass PTR_TO_MEM argument into a function that expects PTR_TO_BTF_ID ?

Sorry, that's poorly phrased. Current kfunc doesn't support PTR_TO_MEM. I meant
it would be allowed now, with the way I implemented things, but there also isn't
a way to signal whether PTR_TO_BTF_ID is expected (hence the question about
bpf_func_proto). I did not understand why that was not done originally (maybe it
was lack of usecase). PTR_TO_CTX works because the type is matched with prog
type, so you can't pass something else there. For other cases the type of
register is considered.

> That sounds like a bug that we need to fix.

--
Kartikeya