Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!uwvax!rutgers!ames!ucbcad!ucbvax!JUPITER.BELLCORE.COM!karn
From: karn@JUPITER.BELLCORE.COM (Phil R. Karn)
Newsgroups: comp.protocols.tcp-ip
Subject: Generalized Subnetting
Message-ID: <8708180007.AA01148@jupiter.planets>
Date: Mon, 17-Aug-87 20:07:53 EDT
Article-I.D.: jupiter.8708180007.AA01148
Posted: Mon Aug 17 20:07:53 1987
Date-Received: Wed, 19-Aug-87 01:31:34 EDT
Sender: daemon@ucbvax.BERKELEY.EDU
Distribution: world
Organization: The ARPA Internet
Lines: 41

The recent discussion about varying subnet sizes was interesting. I have
implemented a "generalized subnetting" scheme in my own IP routing code that
handles this and many similar problems quite well.

Basically, I completely ignore the usual rules that distinguish class A, B
and C addresses. Instead I keep an explicit subnet mask with each entry in
the routing table. (Actually, I use integers between 0 and 32, but the
semantics are the same). When a packet is to be routed, I search the routing
table for the "best" match under control of the mask.  That is, if my
routing table contains

target		width	interface	gateway
128.96.33.101	32	sl0
128.96		16	ec0
0		0	ec0		128.96.41.1

and I present the address 128.96.33.101, all three entries match. In this
case, I select the entry with the greatest width, i.e., the first one.  Note
that the last entry always matches every address; it is the default route.
The first one is a host-specific entry, matching only one IP address.

Note that the routing table points to an INTERFACE name, not a gateway
address. The gateway field is optional. It is filled in only if the target
is to be reached through a gateway on a multiple access subnet, in which
case the gateway field is passed down to the subnet driver for translation
into a subnet address.  If the gateway field is null, the subnet driver is
given the target IP address for translation; this would be done if the
target is directly on the same subnet.  If the subnet is a point-to-point
link, the gateway address field is ignored altogether since the subnet
has no need for link level addresses.

With this scheme you can construct a completely arbitrary internet out of
ad-hoc point-to-point links, incompletely connected broadcast subnets (e.g.,
amateur packet radio), partitioned networks as well as "real" subnets like
Ethernets without filling up your routing table with lots of host-specific
entries. You can "bunch up" address groups that begin with a common bit
pattern, regardless of whether they are really all hosts on the same subnet,
or you can split off one or more hosts that are behind a point-to-point
link, for example. It works very well.

Phil