Re: [PATCH v2] net: don't allow CAP_NET_ADMIN to load non-netdev kernel modules

[Eric Paris <eparis@redhat.com>]

On Thu, 2011-03-24 at 16:57 -0500, Serge E. Hallyn wrote:
> Quoting David Miller (davem@davemloft.net):
> > From: Stephen Hemminger <shemminger@vyatta.com>
> > Date: Thu, 24 Mar 2011 14:39:44 -0700
> > 
> > > This breaks for many of the tunneling protocols, that rely on
> > > autoload for names like "sit0"
> > 
> > Frankly I'm very disappointed in the fallout this has been causing.
> > 
> > Everyone supporting this change, get real, and admit it doing in fact
> > cause a serious regression.
> 
> Sorry, I thought this was causing some extra audit messages but no
> actual breakage?

I've got one report of someone claiming their system broke, but I'm not
convinced I believe it since his dmesg didn't show the magic pr_err()
when it should have.  It's certainly possible this can break someone in
a system which uses fine grained capabilities controls, but I agree it's
pretty unlikely.  My biggest personal concern is that I have a whole
darn bunch of new scary messages which are popping out of people's
computers since they don't have CAP_SYS_MODULE.  While I can silence
them, it's going to hide use of init_module() directly as well, which I
really don't want to hide from the scary logs....

> > If you can't get past that simple fact, you cannot discuss this issue
> > intelligently.
> > 
> > You can't say "userland will fix things up"
> > 
> > Because we're never supposed to break userland in the first place.
> > 
> > There is simply no excuse for this and I want this change reverted
> > both in Linus's tree and in -stable.
> 
> Eric, in this particular case, since we've already done a
> 'capable(CAP_NET_ADMIN)', I woudl argue that doing the check
> for CAP_SYS_ADMIN without auditing failure (even if it requires
> a new helper in capability.c) isn't horrible.  Thoughts?

s/CAP_SYS_ADMIN/CAP_SYS_MODULE/

I can do that.  It was actually my #2 suggestion.  But, I'm certainly
willing to put some of the burden on userspace.  SELinux policy is a
userspace construct and we often force other userspace applications to
fix things they do poorly (even if it gets us a rep for being
'difficult')  Non-SELinux systems aren't going to see this problem,
because basically noone else I know of tries to enforce any kind of
capabilities sets other than all or none, so you'll never see
CAP_NET_ADMIN without CAP_SYS_MODULE.

I guess what it comes down to is that I'm happy to break Fedora user's
with SELinux if in the end it gets us a better system.  I'd be happy to
just rip the whole CAP_SYS_MODULE portion out and blame it on SELinux,
but I know that's not what upstream does.  So given what we have today I
personally would push for a no_audit() interface rather than a complete
revert.   (or maybe a compile option so I can turn off the fallback
altogether and force people to come into compliance)

-Eric