ip_queue_xmit() used illegally

[David Miller <davem@davemloft.net>]

Several users of ip_queue_xmit() use it illegally.

I've only audited L2TP and SCTP so far, and they both cannot use
ip_queue_xmit() with the way they operate currently.

The issue surrounds how the socket binding is maintained in
inet->inet_daddr, inet->inet_saddr etc.

TCP does things right, in that ip_queue_xmit() is only invoked with
inet->inet_daddr and inet->inet_saddr having fully resolved, final,
fully connected values.

This is an absolute requirement because if the socket's route
invalidates (which happens completely asynchronously) it's going to
lookup a new route using whatever is stored in
inet->inet_{daddr,saddr} and then use those addresses to build the
packet.  Even if ->inet_{saddr,daddr} are both zero this will still
emit a packet (bonus points if you know what addresses will be picked,
no peeking at route.c :-).

SCTP stores it's binding information using transports and assosciations
and does not fill in the ->inet_{daddr,saddr} values.

It tries to work around this route issue by checking dst->obsolete
directly in sctp_packet_transmit(), which just makes the race smaller
and does not eliminate it.  ip_queue_xmit() can still end up with
__sk_dst_check() returning NULL and then we end up emitting a
potentially bogus packet.

L2TP supports more of a datagram type socket semantic than a stream
one, it allows unconnected modes of operation.  And for this reason
it also cannot use ip_queue_xmit() legally.

After a quick cursory scan it seem like DCCP is OK.

I think SCTP could potentially be fixed by simply filling in the
inet->inet_{daddr,saddr} values when it makes an internal binding
of the transport via sctp_transport_route().

L2TP on the other hand will need to use another interface to send ipv4
packets because it allows disconnected operation.