[fpc-devel] Re: [fpc-pascal] Website connectivity

Michael Van Canneyt michael at freepascal.org
Mon Aug 17 18:11:18 CEST 2009

On Mon, 17 Aug 2009, Jonas Maebe wrote:

> On 17 Aug 2009, at 17:43, Michael Van Canneyt wrote:
>> The reason for this - which has been given often - was the atypical
>> window size of the router in front of the server: Some ISP's routers or 
>> network configuration could not cope with this atypical window size.
> To nitpick: it's not the window size, but the maximum segment size :)
> And for the people interested in the details: the problems are that router in 
> front of server
> a) blocks incoming ICMP packets that are necessary for automatic MTU 
> detection. As a result, the server cannot detect problems in case a network 
> connection down the line requires a smaller TCP segment size.
> b) set the TCP maximum segment size of all *incoming* packets to 1472. This 
> means that setting the MTU on your computer to a low value (which would then 
> also be used by the server when replying) won't help, since the server always 
> sees an MTU of 1472 bytes.
> c) sets the "don't fragment" flag on outgoing IP packets. This means that 
> even if the server is configured to not add the "don't fragment" flag to 
> packets so they can be fragmented by other routers when sending them over a 
> connection with a smaller MTU, the end result won't change since by the time 
> the packets arrive at those routers, they'll again have a DF flag set.
> In the end, what did work was to enable TCP MTU detection on the server.
> I don't know what kind of router is being used, but I would advise everyone 
> to stay away from it as far as possible. The number of ways in which that 
> thing tries to prevent connectivity with low-MTU endpoints is really 
> astounding.

It's a very common CISCO router :-)

The problem is not really the router itself, but the instructions we got from 
the ISP on how to configure it :(

(totally braindead instructions in my opinion, but arguing didn't really help)


More information about the fpc-devel mailing list