[mod.protocols.tcp-ip] ARP on IBM token ring

leong@ANDREW.CMU.EDU (John Leong) (06/27/86)

For running IP on the IBM token ring, one has to handle ARP quite carefully
(besides boring thing like SAP number as applicable to all LAN's using the 802.2
layer).


The IBM token ring is sort of defined in the official 802.5 spec. What is missing
in that definition is the bridge. Unlike Ethernet, inter-ring bridge is very
much part of the IBM token ring since a basic ring will only support 260 stations
due to potential jitter problem. (Bridge differs from a repeater in that it
will store and forward if necessary). While IBM has contributed the Bridge design
to IEEE, it was not included in the final spec.


The IBM ring bridge works very different from the DEC LANbridge which is totally
transparant to station software. The IBM bridge use source routing. In a simplified
form, when a station does any broadcast, the bridge will forward the broadcast
packet. it will also tag its ID in the Route Information Field of the packet.
The Route Information Field, RIF, is considered as part of the MAC layer and
below the LLC layer. When the broadcast packet gets to the other end, the receiving
station will know the path. (The same thing applies to a special MAC frame called
"resolve"). 


In the case of ARP, the recieving station must store the RIF in its ARP table
since subsequent (non-broadcast) communication will have to be done with source
routing : i.e. including the RIF field in the MAC header. The ARP reply must
also incude the RIF as part of the hardware address so the ARP sender can source
route its packet.


This has at least two implications :


(a) The hardware address of the ARP cache has to be a variable length (subject
to a maximum : there is a maximum length for the RIF field).


(b) There needs to be closer interfacing between the driver and the ARP layer
since the driver must pass up the whole MAC header to the ARP handler so that
the RIF field can be extracted.


Personally, I think DEC's approach makes life much more easier. However, I think
future revision of ARP specification for the 802 LAN's should take the above
into consideration.


John Leong