cannot access a site with kernel 2.6.7 or 2.6.8, 2.6.6 is ok

[Andy Green]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Wednesday 14 July 2004 16:02, Mick Mearns wrote:

>   Yes - that fixed it.
> Although I must admit I have no idea what the heck you are talking
> about ;)

The TCP part of the TCP/IP header on every packet has a place for negotiating 
how much data to keep 'in transit', the 'window'.  Originally TCP set aside 
16 bits in the header to specify window sizes only up to 64K.

In the case of high bandwidth links that have high latency due to the sheer 
distance they travel, 64K isn't enough to allow TCP to saturate the pipe, 
since it takes longer to send back the ACKs for the packets than it takes to 
send 64K.  RFC1323 from May 1992 
(http://www.networksorcery.com/enp/rfc/rfc1323.txt ) is designed to fix this 
by keeping the 16-bit field for window size, but expanding the logical window 
size to 32-bits (allowing 4GB to be in transit(!)) by adding a 0-16 bit 
shifting factor ("Window scaling") for compatability.

The older kernels do not make use of this by default, they keep with a 0 
window scale factor and so can cope with up to 64K window sizes like 
pre-RFC1323 TCP.  But the newer kernels use the scaling factor to allow >64K 
windows, which may be able to make better use of your bandwidth.  The problem 
is some routers out there blow chunks when they see a TCP connection wanting 
>64K window sizes.

Eventually, hopefully, the routers will all get upgraded and we can use the 
better window sizes.  But until that happy day, some sites will grind to a 
halt unless you use the workaround.

- -Andy

- -- 
Automatic actions for USB cameras, cardreaders, memory sticks, MP3 players
http://warmcat.com/usbautocam
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFA9WKyjKeDCxMJCTIRAjQqAJ9ojVnvjYQDxKDFlTigMkBnjsNPrgCfdCUf
jSo1/cDZnsgnQZJyL5lnJXA=
=mlSq
-----END PGP SIGNATURE-----