> >> But quite aside from what it does to the assumptions underlying IP,
> > What does it do to the assumptions underlying IP concretely?
> 
> That every host that can send a datagram into a network has an address
> in that network.  (This is why your suggestion, constructing a
> source-route option, isn't actually a problem, because it doesn't
> involve hosts sending into networks to which they have an interface
> physically connected but in which they do not appear in IP terms.)

I think that even without constructing a source-route option, nothing
would break. In my example, 192.168.2. and 192.168.4. were intended to
be physically distinct networks, connected only by the router with
addresses 192.168.2.21 and 192.168.4.21 . So all this issue would be
internal to the sending host (192.168.2.50) and from the network's point
of view, nothing would change. It would be just a change in the internal
structure of the routing tables. (Admittedly quite useless if
source-routing option is not to be added.)

It's true that this topic doesn't have much common with the rest of this
thread, apart from the word "routing" :-) That's maybe what mislead you.

> 
> >> how would you cut off infinite loops?
> > Isn't the Source Routing option (for IPv4) limited in size?  There
> > would be no point in having more loops than this limit.
> 
> True, and that limit _is_ relatively small.  Okay, I'm answered.

OK, so I'll have to implement it. Expect patches by the end of 
summer :-) .

Thanks for your reply.
Bye	Pavel