In message <19990921181025.F16587@theory.cs.uni-bonn.de>Ignatios Souvatzis writ
es
>On Tue, Sep 21, 1999 at 06:07:39PM +0200, Ignatios Souvatzis wrote:
>> On Tue, Sep 21, 1999 at 08:59:39AM -0700, M Graff wrote:
>> > Bill Sommerfeld <sommerfeld@orchard.arlington.ma.us> writes:
>> > 
>> > > Blocking all ICMP's breaks *all* known Path MTU discovery
>> > > implementations, because PMTUD depends on receiving ICMP
>> > > unreachable/"fragmentation needed but DF set" errors from the
>> > > bottleneck router.
>> > 
>> > I thought we would back off to a smaller packet size if we "lost
>> > contact" after N seconds.  Am I wrong?
>> 
>> No. We try bigger packets after N seconds, and use the next-hop MTU
>> suggested by the ICMP packet too big messages.
>
>well, at least that's what the RFC says. I didn't look at our code recently.

Yah, that's what our code should do.  We should also (IMHO) have
code that will back off the size after some configurable number of
timeouts.  There is some obviously-wrong code #ifdef'ed out in
the timeout handling code that I inserted when I put in PMTUD.
At the time, I hoped to return to it and fix it up soon, but 
I've not yet had the chance.  I'm currently integrating the
PSC SACK code, so I'm not going to have the time to do it right away.

Alas, none of this will help in the reception of web pages, anyway,
since this is a sender-side fix.

Kevin
kml@patheticgeek.net