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

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

...
  On Thu, 13 Oct 2022 06:01:19 +0200
 Stefano Brivio <sbrivio(a)redhat.com> wrote:
 
  On Tue, 11 Oct 2022 16:40:15 +1100
 David Gibson <david(a)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