Re: [PATCH bpf] bpf: fix unused-var without NETDEVICES

From: Alexei Starovoitov
Date: Wed Jun 03 2020 - 14:15:04 EST


On Wed, Jun 03, 2020 at 11:12:01AM +0200, Matthieu Baerts wrote:
> Hi Ferenc,
>
> On 03/06/2020 10:56, Ferenc Fejes wrote:
> > Matthieu Baerts <matthieu.baerts@xxxxxxxxxxxx> ezt Ãrta (idÅpont:
> > 2020. jÃn. 3., Sze, 10:11):
> > >
> > > A recent commit added new variables only used if CONFIG_NETDEVICES is
> > > set.
> >
> > Thank you for noticing and fixed this!
> >
> > > A simple fix is to only declare these variables if the same
> > > condition is valid.
> > >
> > > Other solutions could be to move the code related to SO_BINDTODEVICE
> > > option from _bpf_setsockopt() function to a dedicated one or only
> > > declare these variables in the related "case" section.
> >
> > Yes thats indeed a cleaner way to approach this. I will prepare a fix for that.
>
> I should have maybe added that I didn't take this approach because in the
> rest of the code, I don't see that variables are declared only in a "case"
> section (no "{" ... "}" after "case") and code is generally not moved into a
> dedicated function in these big switch/cases. But maybe it makes sense here
> because of the #ifdef!
> At the end, I took the simple approach because it is for -net.
>
> In other words, I don't know what maintainers would prefer here but I am
> happy to see any another solutions implemented to remove these compiler
> warnings :)

since CONFIG_NETDEVICES doesn't change anything in .h
I think the best is to remove #ifdef CONFIG_NETDEVICES from net/core/filter.c
and rely on sock_bindtoindex() returning ENOPROTOOPT
in the extreme case of oddly configured kernels.