On Wed, 2005-05-25 at 13:16 -0700, David S.Miller wrote:
> From: Kenneth Johansson <ken@kenjo.org>
> Date: Wed, 25 May 2005 22:13:27 +0200
> 
> > Is there some fundamental difference in how nfs over upd and tcp is
> > handled regarding the packet contents like tcp using the tcp checksum
> > and udp not using the udp checksum or something like that?
> > 
> > Are there any counters for checksum errors in udp and tcp that can be
> > read ?? I faild to spot anything in /proc.
> 
> If you are on a gigabit or faster network, IPv4 fragment sequence
> numbers can wrap and if you are very unlucky the checksums will
> match as well corrupting your data.  This is a fatal limitation of
> the small 16-bit IPv4 framgent ID.
> 
> Use TCP for NFS unless you want NFS data corruption.
> 

Unlikely to be the case this time. I get a sequence of 28 bytes that is
wrong in the data and often the wrong data is a copy from data 64 or 128
byte earlier in the file. If this was not on a PC with cache coherency I
would guess that someone forgot to do a cache invalidate/flush. But I do
wonder why I only see this problem with nfs over udp.