[PATCH] selinux: Fix error priority for bind with AF_UNSPEC on AF_INET6 socket

[Mickaël Salaün]

(Removing Alexey Kodanev because the related address is no longer
valid.)

On Thu, Dec 28, 2023 at 07:19:07PM -0500, Paul Moore wrote:
> On Thu, Dec 28, 2023 at 6:39 AM Mickaël Salaün <mic at digikod.net> wrote:
> >
> > The IPv6 network stack first checks the sockaddr length (-EINVAL error)
> > before checking the family (-EAFNOSUPPORT error).
> >
> > This was discovered thanks to commit a549d055a22e ("selftests/landlock:
> > Add network tests").
> >
> > Cc: Alexey Kodanev <alexey.kodanev at oracle.com>
> > Cc: Eric Paris <eparis at parisplace.org>
> > Cc: Konstantin Meskhidze <konstantin.meskhidze at huawei.com>
> > Cc: Paul Moore <paul at paul-moore.com>
> > Cc: Stephen Smalley <stephen.smalley.work at gmail.com>
> > Reported-by: Muhammad Usama Anjum <usama.anjum at collabora.com>
> > Closes: https://lore.kernel.org/r/0584f91c-537c-4188-9e4f-04f192565667@collabora.com
> > Fixes: 0f8db8cc73df ("selinux: add AF_UNSPEC and INADDR_ANY checks to selinux_socket_bind()")
> > Signed-off-by: Mickaël Salaün <mic at digikod.net>
> > ---
> >  security/selinux/hooks.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> >
> > diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c
> > index feda711c6b7b..9fc55973d765 100644
> > --- a/security/selinux/hooks.c
> > +++ b/security/selinux/hooks.c
> > @@ -4667,6 +4667,10 @@ static int selinux_socket_bind(struct socket *sock, struct sockaddr *address, in
> >                                 return -EINVAL;
> >                         addr4 = (struct sockaddr_in *)address;
> >                         if (family_sa == AF_UNSPEC) {
> > +                               if (sock->sk->__sk_common.skc_family ==
> > +                                           AF_INET6 &&
> > +                                   addrlen < SIN6_LEN_RFC2133)
> > +                                       return -EINVAL;
> 
> Please use sock->sk_family to simplify the conditional above, or
> better yet, use the local variable @family as it is set to the sock's
> address family near the top of selinux_socket_bind()

Correct, I'll send a v2 with that.

> ... although, as
> I'm looking at the existing code, is this patch necessary?
> 
> At the top of the AF_UNSPEC/AF_INET case there is an address length check:
> 
>   if (addrlen < sizeof(struct sockaddr_in))
>     return -EINVAL;

This code is correct but not enough in the case of an IPv6 socket.

> 
> ... which I believe should be performing the required sockaddr length
> check (and it is checking for IPv4 address lengths not IPv6 as in the
> patch).  I see that we have a similar check for AF_INET6, so we should
> be covered there as well.

The existing similar check (addrlen < SIN6_LEN_RFC2133) is when the
af_family is AF_INET6, but this patch adds a check for AF_UNSPEC on an
PF_INET6 socket. The IPv6 network stack first checks that the addrlen is
valid for an IPv6 address even if the requested af_family is AF_UNSPEC,
hence this patch.

> 
> I'm probably still in a bit of a holiday fog, can you help me see what
> I'm missing here?

The tricky part is that AF_UNSPEC can be checked against the PF_INET or
the PF_INET6 socket implementations, and the return error code may not
be the same according to addrlen, especially when
sizeof(struct sockaddr_in) < addrlen < SIN6_LEN_RFC2133

The (new) Landlock network tests check this kind of corner case to make
sure the same error codes are return with and without a Landlock
sandbox. Muhammad reported that some of these tests failed on KernelCI
and I found that, when SELinux is enabled (which is the case with the
defconfig), SElinux gets the request after Landlock and returns a wrong
error code (before the network stack can do anything).
See tools/testing/selftests/landlock/net_test.c +728
which checks with and without a Landlock sandbox.

I tested this patch with SELinux and Landlock enabled, and all the
Landlock tests pass.

I'm working on a more global approach to cover all LSMs, with more
checks and Landlock tests, but this will be more complex and then will
take more time to review.