> That is, is PMTU discovery the one and only application, ever, of
> using the DF flag?

I don't think so.  I can conceive of embedded boxes which would not
want to have to deal with IP fragmentation and reassembly..  (not so
much from the code size, but from issues relating to the memory*time
footprint caused by fragmented packets where some of the fragments had
been lost..).

That said, it's OK to fragment a packet with DF set at the link level
as long as it's reassembled before it is forwarded by the router on
the other end of the link.  But clearing the DF bit and fragmenting at
the IP layer is wrong.

Tunnels make this more interesting, since the "link layer" is another
instance of IP...

					- Bill