Merge branch 'inet_fwmark_reflect'
Lorenzo Colitti says:
====================
Make mark-based routing work better with multiple separate networks.
Mark-based routing (ip rule fwmark 17 lookup 100) combined with
either iptables marking (iptables -j MARK --set-mark 17) or
application-based marking (the SO_MARK setsockopt) are a good
way to deal with connecting simultaneously to multiple networks.
Each network can be given a routing table, and ip rules can
be configured to make different fwmarks select different
networks. Applications can select networks them by setting
appropriate socket marks, and iptables rules can be used to
handle non-aware applications, enforce policy, etc.
This patch series improves functionality when mark-based routing
is used in this way. Current behaviour has the following
limitations:
1. Kernel-originated replies that are not associated with a
socket always use a mark of zero. This means that, for
example, when the kernel sends a ping reply or a TCP reset,
it does not send it on the network from which it received the
original packet.
2. Path MTU discovery, which is triggered by incoming packets,
does not always work correctly, because the routing lookups it
uses to clone routes do not take the fwmark into account and
thus can happen in the wrong routing table.
3. Application-based marking works well for outbound connections,
but does not work well for incoming connections. Marking a
listening socket causes that socket to only accept
connections from a given network, and sockets that are
returned by accept() are not marked (and are thus not routed
correctly).
sysctl. This causes route lookups for kernel-generated replies
and PMTUD to use the fwmark of the packet that caused them.
which causes TCP sockets returned by accept() to be marked with
the same mark that sent the intial SYN packet.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>