On 1/8/2025 12:00 AM, Paul Moore wrote:

On Tue, Jan 7, 2025 at 3:16 PM Stephen Smalley
[off-list ref] wrote:

quoted

On Fri, Dec 13, 2024 at 3:09 PM Paul Moore [off-list ref] wrote:

quoted

On Fri, Dec 13, 2024 at 11:40 AM Mikhail Ivanov
[off-list ref] wrote:

quoted

On 12/13/2024 6:46 PM, Stephen Smalley wrote:

quoted

On Fri, Dec 13, 2024 at 5:57 AM Mikhail Ivanov
[off-list ref] wrote:

quoted

On 12/12/2024 8:50 PM, Mickaël Salaün wrote:

quoted

This looks good be there are other places using sk->sk_family that
should also be fixed.

Thanks for checking this!

For selinux this should be enough, I haven't found any other places
where sk->sk_family could be read from an IPv6 socket without locking.

I also would like to prepare such fix for other LSMs (apparmor, smack,
tomoyo) (in separate patches).

I'm wondering about the implications for SELinux beyond just
sk->sk_family access, e.g. SELinux maps the (family, type, protocol)
triple to a security class at socket creation time via
socket_type_to_security_class() and caches the security class in the
inode_security_struct and sk_security_struct for later use.

IPv6 and IPv4 TCP sockets are mapped to the same SECCLASS_TCP_SOCKET
security class. AFAICS there is no other places that can be affected by
the IPV6_ADDFORM transformation.

Yes, thankfully we don't really encode the IP address family in any of
the SELinux object classes so that shouldn't be an issue.  I also
don't think we have to worry about the per-packet labeling protocols
as it's too late in the communication to change the socket's
associated packet labeling, it's either working or it isn't; we should
handle the mapped IPv4 address already.

I am a little concerned about bind being the only place where we have
to worry about accessing sk_family while the socket isn't locked.  As
an example, I'm a little concerned about the netfilter code paths; I
haven't chased them down, but my guess is that the associated
socket/sock isn't locked in those cases (in the relevant output and
postroute cases, forward should be a non-issue).

We still need an answer on this.

Sorry for the late reply,
I found out that security_sock_rcv_skb() can also be called without
locking the IPv6 socket (this can be easily verified by manual testing).

Netfilter hooks seems to be ok, family value is taken from the
nf_hook_state structure, so there is no access to sk->sk_family.

SCTP and MPTCP hooks should not be considered, because IPV6_ADDRFORM is
only available for TCP, UDP and UDPLITE protocols.

There are 2 more functions that access sk_family:
* security_sock_graft() - socket is locked by inet_accept(),
* security_inet_conn_established() - socket is locked by connect(2) or
   in BH context (Cf. tcp_v6_rcv).

quoted

quoted

How bad is the performance impact of READ_ONCE()?  In other words, how
stupid would it be to simply do all of our sock->sk_family lookups
using READ_ONCE()?

I could be wrong, but I don't think there is any overhead except on Dec Alpha.

Then perhaps the right answer is to use it everywhere.

Indeed, using READ_ONCE() in the considered hooks should not lead to
any overhead. I wonder if it would be better not to touch the SCTP
and MPTCP hooks anyway. Adding READ_ONCE() in selinux_sock_graft() is
fine if you think it's better this way.