Streams and multiplexing

dmr at dutoit.UUCP dmr at dutoit.UUCP
Tue Oct 8 16:14:48 AEST 1985


Steven J. Langdon claimed that without multiplexing one couldn't
do a proper kernel-resident version of TCP/IP in the V8 stream context.
Here's how it's done.

It is still true in our system that stream multiplexing does not occur,
in the sense that every stream connection has (from the point of view
of the formal data structures) exactly two ends, one at a user process,
and the other at a device or another process.  However, this has, in
practice, not turned out to be a problem.  Say you have a hardware
device that hands you packets with a channel (or socket) number buried
inside in some complicated format.  The general scheme to handle the
situation uses both a line discipline (stream filter module) and
associated code that, to the system, looks like a stream device driver
with several minor devices; these have entries in /dev.

A watchdog process opens the underlying real device, and pushes
the stream module.   Arriving packets from the real device
are passed to this module, where they are analyzed,
and then given to the appropriate associated pseudo-device.
Likewise, messages written on the pseudo-device are shunted over to
the line discipline, where they are encoded appropriately and sent
to the real device.  This is where the multiplexing-demultiplexing
occurs; formally, it is outside of the stream structure, because
the data-passing doesn't follow the forward and backward links
of the stream modules.  However, the interfaces of the combined
larger module obey stream rules.

For example, IP works this way:  The IP line discipline looks at the
type field of data arriving from the device, and determines whether the
packet is TCP or UDP or ARP or whatever, and shunts it off to the
stream associated with /dev/ip6 or /dev/ip17 or whatever the numbers
are.

TCP, of course, is multiplexed as well.  So there is a TCP line
discipline, and a bunch of TCP devices; a watchdog process opens
/dev/ip6, and pushes the TCP line discipline; then the TCP packets it
gets are parcelled out to the appropriate /dev/tcpXX device.  Each TCP
device looks like the end of a stream, and may, of course, have other
modules (e.g.  tty processor) inserted in this stream.

UDP sits on top of IP in the same way.

This example is complicated, because (TCP,UDP)/IP is.  However, it
works well.  In particular, the underlying real device can be either an
ethernet or our own Datakit network; the software doesn't care.  For
example, from my machine, I can type "rlogin purdy" and connect to a
Sequent machine running 4.2; the TCP connection goes over Datakit to
machine "research" where it is gatewayed to a local ethernet that purdy
is connected to.

A further generalization (that we haven't made) is in principle easy:
there can be protocol suites other than IP on an Ethernet cable.  So
there could be another layer to separate IP from XNS from Chaosnet, etc.

	Dennis Ritchie



More information about the Comp.unix.wizards mailing list