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

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

On Sat, Nov 06, 2021 at 02:19:08AM IST, Alexei Starovoitov wrote:
> On Thu, Nov 04, 2021 at 06:25:03PM +0530, Kumar Kartikeya Dwivedi wrote:
> > 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.
>
> So it's just to special case r0=PTR_TO_BTF_ID_OR_NULL instead of
> PTR_TO_BTF_ID that it's doing by default now?
> Then could you use a btf_id list to whitelist all such funcs that needs _OR_NULL
> variant and just do a search in that list in check_kfunc_call() ?
> Instead of adding get_kfunc_return_type() callback.
>

Hm, good idea, that should work for now.

> > 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.
>
> No need to hack into bpf_func_proto. All kernel funcs have BTF. It's all we need.
> The _OR_NULL part we will eventually be able to express with btf_tag when
> it's supported by both gcc and clang.
>

More on this below.

> > > > 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.
>
> Right. btf_check_kfunc_arg_match doesn't allow ptr_to_mem yet.
> There is no signalling needed.
> All args passed by the program into kfunc have to be either exact
> PTR_TO_BTF_ID or conversions from PTR_TO_SOCK*.
>

I should have been clearer again :). Sorry for that.

Right now only PTR_TO_BTF_ID and PTR_TO_SOCK and scalars are supported, as you
noted, for kfunc arguments.

So in 3/6 I move the PTR_TO_CTX block before btf_is_kernel check, that means if
reg type is PTR_TO_CTX and it matches the argument for the program, it will use
that, otherwise it moves to btf_is_kernel(btf) block, which checks if reg->type
is PTR_TO_BTF_ID or one of PTR_TO_SOCK* and does struct match for those. Next, I
punt to ptr_to_mem for the rest of the cases, which I think is problematic,
since now you may pass PTR_TO_MEM where some kfunc wants a PTR_TO_BTF_ID.

But without bpf_func_proto, I am not sure we can decide what is expected in the
kfunc. For something like bpf_sock_tuple, we'd want a PTR_TO_MEM, but taking in
a PTR_TO_BTF_ID also isn't problematic since it is just data, but for a struct
embedding pointers or other cases, it may be a problem.

For PTR_TO_CTX in kfunc case, based on my reading and testing, it will reject
any attempts to pass anything other than PTR_TO_CTX due to btf_get_prog_ctx_type
for that argument. So that works fine.

To me it seems like extending with some limited argument checking is necessary,
either using tagging as you mentioned or bpf_func_proto, or some other hardcoded
checking for now since the number of helpers needing this support is low.

Please correct me if I'm wrong about any of this.

> Passing rX=PTR_TO_CTX into kfunc should not work. If I'm reading the code
> correctly it's not allowed. I'm not sure why you're saying it can be done.
> It's possible to pass PTR_TO_CTX into another bpf prog's global function.
> The same btf_check_func_arg_match() helper checks both cases (global funcs and kfuncs).
> Maybe that's where the confusion comes from?
>
> Same with if (ptr_to_mem_ok). It's only for passing PTR_TO_MEM
> into bpf prog's global function.
> We can extend the verifier and allow PTR_TO_MEM into kfunc that
> has 'long *' prototype, for example.
> But it doesn't sound like the use case you have in mind.

--
Kartikeya