On Sat, Apr 27, 2019 at 8:39 PM Jakub Kicinski
<jakub.kicinski@xxxxxxxxxxxxx> wrote:
>
> On Sat, 27 Apr 2019 12:48:25 +0200, Alban Crequy wrote:
> > On Fri, Apr 26, 2019 at 11:03 PM Jakub Kicinski
> > <jakub.kicinski@xxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, 26 Apr 2019 17:48:45 +0200, Alban Crequy wrote:
> > > > In the unlikely case where network namespaces are not compiled in
> > > > (CONFIG_NET_NS=n), the verifier will not allow access to ->netns_*.
> > >
> > > Naive question - why return an error? init_net should always be there,
> > > no?
> >
> > True for netns_dev. However, without CONFIG_NET_NS, we cannot access netns_ino:
> >
> > (struct sock_common).possible_net_t.(struct net *):
> >
> > typedef struct {
> > #ifdef CONFIG_NET_NS
> > struct net *net;
> > #endif
> > } possible_net_t;
> >
> > And I don't think it would make much sense to allow access to
> > netns_dev but not netns_ino.
>
> Right, if CONFIG_NET_NS=n we could just take the pointer to init_net
> directly, and not worry about the field. IMHO it'd be preferable to
> changing the UAPI based on kernel config, but I don't feel super
> strongly.
I see the point about not changing the UAPI. So I will update the patch to:
- return netns_dev unconditionally, regardless of CONFIG_NET_NS
- return netns_ino with either the correct value or zero depending on
CONFIG_NET_NS.
