Greg A. Woods writes:
> No memory necessary, well, not very much anyway.  If you get a packet
> with the DF *and* you have to return an ICMP "needs frag" reply for it,
> *and* the packet is part of an already established TCP circuit, then you
> record this instance in a small cache of such things.  If you get a
> second packet of the same size and have to do the same again then you
> increment a counter in the first record and mark the time of the second
> event.  When the N'th identical sized packet comes through from the same
> source then you set a flag in the cache item, turn off the DF bit, and
> pass the packet on (in fragments).  The flag says to ignore the DF bit
> on all packets from that source address.

This totally violates the required behavior of routers. It is a
bad idea. Among other things, it probably breaks path MTU
discovery. I'd very strongly suggest forgeting about this.

Perry