tcp: apply CLOEXEC to accepted socket, not listener

The accept path calls set_cloexec(sd) after accept(). That re-flags the
listening socket, which is already CLOEXEC from create_socket_tcp(), and
leaves new_sd inheritable. As a result, client-connect and auth scripts
spawned after accept can inherit the connected socket and read or write
the raw TCP stream. This defeats the stated intent to prevent scripts from
accessing the client socket.

This bug was found using ZeroPath.

Signed-off-by: Joshua Rogers <MegaManSec@users.noreply.github.com>
Acked-by: Gert Doering <gert@greenie.muc.de>
Message-Id: <-MNw5Hu8h0rHV18x36ISt7V0UHchIO4i-JoAeV_wlxS1AmDIAe7YVYNput3_r2hiu3HhwxkhGyUhv4-iH_E7mf7nGjvocmGXlDq7Tjly5cE=@joshua.hu>
URL: https://www.mail-archive.com/openvpn-devel@lists.sourceforge.net/msg33823.html
Signed-off-by: Gert Doering <gert@greenie.muc.de>
(cherry picked from commit c0d96fd873)

src/openvpn/socket.c

@@ -1270,7 +1270,7 @@ socket_do_accept(socket_descriptor_t sd,
     {
         /* set socket file descriptor to not pass across execs, so that
          * scripts don't have access to it */
-        set_cloexec(sd);
+        set_cloexec(new_sd);
     }
     return new_sd;
 }