On Mon, Mar 30, 2020 at 5:38 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Mon, Mar 30, 2020 at 05:05:13PM -0700, Andrey Ignatov wrote:
> > >
> > > +#define BPF_LINK_CREATE_LAST_FIELD link_create.flags
> > > +static int link_create(union bpf_attr *attr)
> > > +{
> >
> > From what I see this function does not check any capability whether the
> > existing bpf_prog_attach() checks for CAP_NET_ADMIN.
>
> Great catch! It's a bug.
> I fixed it up.
Thanks!
>
> > This is pretty importnant difference but I don't see it clarified in the
> > commit message or discussed (or I missed it?).
Yeah, not intentional, thanks for catching!
> >
> > Having a way to attach cgroup bpf prog by non-priv users is actually
> > helpful in some use-cases, e.g. systemd required patching in the past to
> > make it work with user (non-priv) sessions, see [0].
> >
> > But in other cases it's also useful to limit the ability to attach
> > programs to a cgroup while using bpf_link so that only the thing that
> > controls cgroup setup can attach but not any non-priv process running in
> > that cgroup. How is this use-case covered in BPF_LINK_CREATE?
> >
> >
> > [0] https://github.com/systemd/systemd/pull/12745
>
> yeah. we need to resurrect the discussion around CAP_BPF.
>
> PS
> pls trim your replies.
