Date:   Thu, 9 May 2019 09:17:28 -0400
From:   Paul Moore <paul@...l-moore.com>
To:     Paolo Abeni <pabeni@...hat.com>
Cc:     davem@...emloft.net, selinux@...r.kernel.org,
        netdev@...r.kernel.org, Tom Deseyn <tdeseyn@...hat.com>,
        Richard Haines <richard_c_haines@...nternet.com>,
        Stephen Smalley <sds@...ho.nsa.gov>,
        Marcelo Ricardo Leitner <marcelo.leitner@...il.com>
Subject: Re: [PATCH net] selinux: do not report error on connect(AF_UNSPEC)

On Thu, May 9, 2019 at 4:40 AM Paolo Abeni <pabeni@...hat.com> wrote:
> On Wed, 2019-05-08 at 17:17 -0400, Paul Moore wrote:
> > On Wed, May 8, 2019 at 2:55 PM Stephen Smalley <sds@...ho.nsa.gov> wrote:
> > > On 5/8/19 2:27 PM, Marcelo Ricardo Leitner wrote:
> > > > On Wed, May 08, 2019 at 02:13:17PM -0400, Stephen Smalley wrote:
> > > > > On 5/8/19 2:12 PM, Stephen Smalley wrote:
> > > > > > On 5/8/19 9:32 AM, Paolo Abeni wrote:
> > > > > > > calling connect(AF_UNSPEC) on an already connected TCP socket is an
> > > > > > > established way to disconnect() such socket. After commit 68741a8adab9
> > > > > > > ("selinux: Fix ltp test connect-syscall failure") it no longer works
> > > > > > > and, in the above scenario connect() fails with EAFNOSUPPORT.
> > > > > > >
> > > > > > > Fix the above falling back to the generic/old code when the address
> > > > > > > family
> > > > > > > is not AF_INET{4,6}, but leave the SCTP code path untouched, as it has
> > > > > > > specific constraints.
> > > > > > >
> > > > > > > Fixes: 68741a8adab9 ("selinux: Fix ltp test connect-syscall failure")
> > > > > > > Reported-by: Tom Deseyn <tdeseyn@...hat.com>
> > > > > > > Signed-off-by: Paolo Abeni <pabeni@...hat.com>
> > > > > > > ---
> > > > > > >    security/selinux/hooks.c | 8 ++++----
> > > > > > >    1 file changed, 4 insertions(+), 4 deletions(-)
> > > > > > >
> > > > > > > diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c
> > > > > > > index c61787b15f27..d82b87c16b0a 100644
> > > > > > > --- a/security/selinux/hooks.c
> > > > > > > +++ b/security/selinux/hooks.c
> > > > > > > @@ -4649,7 +4649,7 @@ static int
> > > > > > > selinux_socket_connect_helper(struct socket *sock,
> > > > > > >            struct lsm_network_audit net = {0,};
> > > > > > >            struct sockaddr_in *addr4 = NULL;
> > > > > > >            struct sockaddr_in6 *addr6 = NULL;
> > > > > > > -        unsigned short snum;
> > > > > > > +        unsigned short snum = 0;
> > > > > > >            u32 sid, perm;
> > > > > > >            /* sctp_connectx(3) calls via selinux_sctp_bind_connect()
> > > > > > > @@ -4674,12 +4674,12 @@ static int
> > > > > > > selinux_socket_connect_helper(struct socket *sock,
> > > > > > >                break;
> > > > > > >            default:
> > > > > > >                /* Note that SCTP services expect -EINVAL, whereas
> > > > > > > -             * others expect -EAFNOSUPPORT.
> > > > > > > +             * others must handle this at the protocol level:
> > > > > > > +             * connect(AF_UNSPEC) on a connected socket is
> > > > > > > +             * a documented way disconnect the socket.
> > > > > > >                 */
> > > > > > >                if (sksec->sclass == SECCLASS_SCTP_SOCKET)
> > > > > > >                    return -EINVAL;
> > > > > > > -            else
> > > > > > > -                return -EAFNOSUPPORT;
> > > > > >
> > > > > > I think we need to return 0 here.  Otherwise, we'll fall through with an
> > > > > > uninitialized snum, triggering a random/bogus permission check.
> > > > >
> > > > > Sorry, I see that you initialize snum above.  Nonetheless, I think the
> > > > > correct behavior here is to skip the check since this is a disconnect, not a
> > > > > connect.
> > > >
> > > > Skipping the check would make it less controllable. So should it
> > > > somehow re-use shutdown() stuff? It gets very confusing, and after
> > > > all, it still is, in essence, a connect() syscall.
> > >
> > > The function checks CONNECT permission on entry, before reaching this
> > > point.  This logic is only in preparation for a further check
> > > (NAME_CONNECT) on the port.  In this case, there is no further check to
> > > perform and we can just return.
> >
> > I agree with Stephen, in the connect(AF_UNSPEC) case the right thing
> > to do is to simply return with no error.
>
> The 'default:' case is catching any address family other than
> INET{4,6}, but I guess you argument still applies - selinux should not
> do name check for unknown protocols ?!?

If the code doesn't understand how to parse the port/"name" info it
can't really do a useful name_connect check, this is why we return
-EAFNOSUPPORT in the default case (or -EINVAL in the case of SCTP).
However, the connect/AF_UNSPEC case is a bit of a special case and as
such I probably needs special handling.

My initial thinking is that we should do the AF_UNSPEC check
immediately after the sock_has_perm() check in
selinux_socket_connect_helper():

       err = sock_has_perm(sk, SOCKET__CONNECT);
       if (err)
               return err;
       if (addrlen < offsetofend(struct sockaddr, sa_family))
               return -EINVAL;
       if (address->sa_family == AF_UNSPEC)
               return 0;

... we can then remove the addrlen check from inside the TCP/DCCP/SCTP
if-true block later in the function.  There is the downside the we are
now adding some additional code that executes for each connect() call
(as opposed to just TCP/DCCP/SCTP), but this seems much cleaner from a
conceptual point of view and I expect the overhead to be in the
"unmeasurable" range.

> > I would also suggest that since this patch only touches the SELinux
> > code it really should go in via the SELinux tree and not netdev; this
> > will help avoid merge conflicts in the linux-next tree and during the
> > merge window.  I think the right thing to do at this point is to
> > create a revert patch (or have DaveM do it, I'm not sure what he
> > prefers in situations like this) for this commit, make the adjustments
> > that Stephen mentioned and submit them for the SELinux tree.
>
> Sorry, my fault, I sent the email to both MLs for more awareness, I
> should have used a different subject prefix.

It's not a big deal for patches this small, but since you're going to
respin this patch anyway I figured it would be worth mentioning.
Also, I have no object to posting to multiple MLs when appropriate (it
seems appropriate here); I think the problem here was the "[PATCH
net]" which caused DaveM to pull it into his tree.

> @DaveM: if it's ok for you, I'll send a revert for this on netdev and
> I'll send a v2 via the selinux ML, please let me know!

-- 
paul moore
www.paul-moore.com

Powered by blists - more mailing lists