Hi, Neil.
Neil Brown wrote:
>
> On Wednesday April 5, [EMAIL PROTECTED] wrote:
> > [ Is there a Linux NFS mailing list? ]
> >
> > I'm discussing a network driver problem with a gentleman in Germany who
> > says that his NFS client's driver problem can cause incorrect data to
> > end up on the server's disks; "files written with a lot of @@@@@@@... in
> > their body".
> >
> > I'm trying to work out how this can happen. Does NFS not use UDP
> > checksumming? If so, is this wise?
>
> [EMAIL PROTECTED] (I think).
Ah. I looked at nfs.sourceforge.net, but it looke a litle derelict.
> At they '@' (character 64) or '^@' (character 0).
Not sure.
> I have heard NFS described as
> Nuls Frequently Substituted
>
> but this only happens when you use "soft" mounts. Soft mounts are a
> bad idea, especially when you use programs that don't check for read
> errors properly.
> Could this be the cause? Try "hard" mounts and see if you can
> reproduce it.
Could be. These people are using Linux <-> Linux in a Beowulf cluster.
But I realise I've been silly - I personally experienced this problem
four weeks ago.
I had a 2.3.49 NFS server and a 2.2.13 client. The server had a NIC
problem. Receive packets were occasionally getting a few bits flipped
as they were being DMA'ed into RAM (the EISA clock was too high).
These bad bits were making it all the way onto the disk, which surprises
me. They will have had a good MAC-layer checksum but surely they must
have had a bad UDP checksum.
I just had an ethereal peek at the NFS traffic between a 2.3.99 client
and a netapp server. The checksums are there, and they're not 0xffff.
I have a feeling I chose the wrong mailing list...
--
-akpm-
