Re: [PATCH 07/10] isolation: Replace drop_caps() with a version that actually does something

[Stefano Brivio <sbrivio@redhat.com>]

On Thu, 13 Oct 2022 15:08:02 +0200
Stefano Brivio <sbrivio@redhat.com> wrote:

> On Thu, 13 Oct 2022 06:01:19 +0200
> Stefano Brivio <sbrivio@redhat.com> wrote:
> 
> > On Tue, 11 Oct 2022 16:40:15 +1100
> > David Gibson <david@gibson.dropbear.id.au> wrote:
> >   
> > > @@ -251,7 +275,19 @@ int isolate_prefork(struct ctx *c)
> > >  		return -errno;
> > >  	}
> > >  
> > > -	drop_caps();	/* Relative to the new user namespace this time. */
> > > +	/* Drop capabilites in our new userns */
> > > +	if (c->mode == MODE_PASTA) {
> > > +		/* Keep CAP_SYS_ADMIN, so that we can setns() to the
> > > +		 * netns when we need to act upon it
> > > +		 */
> > > +		ns_caps |= 1UL << CAP_SYS_ADMIN;
> > > +		/* Keep CAP_NET_BIND_SERVICE, so we can splice
> > > +		 * outbound connections to low port numbers
> > > +		 */
> > > +		ns_caps |= 1UL << CAP_NET_BIND_SERVICE;
> > > +	}
> > > +
> > > +	drop_caps_ep_except(ns_caps);    
> > 
> > Hmm, I didn't really look into this yet, but there seems to be an issue
> > with filesystem-bound network namespaces now. Running something like:
> > 
> >   pasta --config-net --netns /run/user/1000/netns/netns-6466ff4b-1efc-2b58-685b-cbc12feb9ccc
> > 
> > (from Podman), this happens:
> > 
> > [...]
> >
> > [pid 1763223] setns(7, CLONE_NEWNET)    = -1 EPERM (Operation not permitted)  
> 
> Ah, "of course". Podman calls us with UID 0 in the user namespace it
> just created, so if we drop CAP_SYS_ADMIN in isolate_initial() we can't
> join the network namespace, and if we drop CAP_NET_ADMIN we can't
> configure it.
> 
> So for that case (and only for that, I suppose), we need something like
> (tested):
> 
> diff --git a/isolation.c b/isolation.c
> index 1769180..fee6dbd 100644
> --- a/isolation.c
> +++ b/isolation.c
> @@ -190,7 +190,7 @@ void isolate_initial(void)
>          * namespace if we have it, so that we can forward low ports
>          * into the guest/namespace
>          */
> -       drop_caps_ep_except((1UL << CAP_NET_BIND_SERVICE));
> +       drop_caps_ep_except(BIT(CAP_SYS_ADMIN) | BIT(CAP_NET_ADMIN));
>  }
> 
> ...which is a bit pointless. Better than *any* capability, but not by
> far.
> 
> So, if we make this totally independent from configuration, we need
> those two capabilities.
> 
> We could add a "postconf" stage and cover a tiny bit more of conf.c.
> 
> Or we could have a special path in isolate_initial() for the case we
> know we're not in the init namespace.
> 
> I'm not sure. If you have a specific preference/strong opinion I would
> actually be happier. :)

Further on, if we are started as root, we'll fail to drop to 'nobody'
or any other user, if we lose CAP_SETUID and CAP_SETGID here. I have
tested this version of isolate_initial():

	drop_caps_ep_except(BIT(CAP_SYS_ADMIN) | BIT(CAP_NET_ADMIN) |
			    BIT(CAP_SETGID)    | BIT(CAP_SETUID)    |
			    BIT(CAP_NET_BIND_SERVICE));

for any use case I can reasonably think of. Yes, it's a lot -- we
should make it really clear that those are not the capabilities we
actually use at "run time".

-- 
Stefano