Netfilter's flowtable infrastructure
====================================

This documentation describes the software flowtable infrastructure available in
Netfilter since Linux kernel 4.16.

Overview
--------

Initial packets follow the classic forwarding path, once the flow enters the
established state according to the conntrack semantics (ie. we have seen traffic
in both directions), then you can decide to offload the flow to the flowtable
from the forward chain via the 'flow offload' action available in nftables.

Packets that find an entry in the flowtable (ie. flowtable hit) are sent to the
output netdevice via neigh_xmit(), hence, they bypass the classic forwarding
path (the visible effect is that you do not see these packets from any of the
netfilter hooks coming after the ingress). In case of flowtable miss, the packet
follows the classic forward path.

The flowtable uses a resizable hashtable, lookups are based on the following
7-tuple selectors: source, destination, layer 3 and layer 4 protocols, source
and destination ports and the input interface (useful in case there are several
conntrack zones in place).

Flowtables are populated via the 'flow offload' nftables action, so the user can
selectively specify what flows are placed into the flow table. Hence, packets
follow the classic forwarding path unless the user explicitly instruct packets
to use this new alternative forwarding path via nftables policy.

This is represented in Fig.1, which describes the classic forwarding path
including the Netfilter hooks and the flowtable fastpath bypass.

                                         userspace process
                                          ^              |
                                          |              |
                                     _____|____     ____\/___
                                    /          \   /         \
                                    |   input   |  |  output  |
                                    \__________/   \_________/
                                         ^               |
                                         |               |
      _________      __________      ---------     _____\/_____
     /         \    /          \     |Routing |   /            \
  -->  ingress  ---> prerouting ---> |decision|   | postrouting |--> neigh_xmit
     \_________/    \__________/     ----------   \____________/          ^
       |      ^                          |               ^                |
   flowtable  |                     ____\/___            |                |
       |      |                    /         \           |                |
    __\/___   |                    | forward |------------                |
    |-----|   |                    \_________/                            |
    |-----|   |                 'flow offload' rule                       |
    |-----|   |                   adds entry to                           |
    |_____|   |                     flowtable                             |
       |      |                                                           |
      / \     |                                                           |
     /hit\_no_|                                                           |
     \ ? /                                                                |
      \ /                                                                 |
       |__yes_________________fastpath bypass ____________________________|

               Fig.1 Netfilter hooks and flowtable interactions

The flowtable entry also stores the NAT configuration, so all packets are
mangled according to the NAT policy that matches the initial packets that went
through the classic forwarding path. The TTL is decremented before calling
neigh_xmit(). Fragmented traffic is passed up to follow the classic forwarding
path given that the transport selectors are missing, therefore flowtable lookup
is not possible.

Example configuration
---------------------

Enabling the flowtable bypass is relatively easy, you only need to create a
flowtable and add one rule to your forward chain.

        table inet x {
		flowtable f {
			hook ingress priority 0; devices = { eth0, eth1 };
		}
                chain y {
                        type filter hook forward priority 0; policy accept;
                        ip protocol tcp flow offload @f
                        counter packets 0 bytes 0
                }
        }

This example adds the flowtable 'f' to the ingress hook of the eth0 and eth1
netdevices. You can create as many flowtables as you want in case you need to
perform resource partitioning. The flowtable priority defines the order in which
hooks are run in the pipeline, this is convenient in case you already have a
nftables ingress chain (make sure the flowtable priority is smaller than the
nftables ingress chain hence the flowtable runs before in the pipeline).

The 'flow offload' action from the forward chain 'y' adds an entry to the
flowtable for the TCP syn-ack packet coming in the reply direction. Once the
flow is offloaded, you will observe that the counter rule in the example above
does not get updated for the packets that are being forwarded through the
forwarding bypass.

More reading
------------

This documentation is based on the LWN.net articles [1][2]. Rafal Milecki also
made a very complete and comprehensive summary called "A state of network
acceleration" that describes how things were before this infrastructure was
mailined [3] and it also makes a rough summary of this work [4].

[1] https://lwn.net/Articles/738214/
[2] https://lwn.net/Articles/742164/
[3] http://lists.infradead.org/pipermail/lede-dev/2018-January/010830.html
[4] http://lists.infradead.org/pipermail/lede-dev/2018-January/010829.html