[OpenAFS-devel] Client corruption on Linux/AMD64

[Jean-Marc Saffroy]

Hello,

I observe file corruption from a Linux/AMD64 client. If I compare a local 
copy with the file I read on AFS, I see different checksums:

$ md5sum ImpactofDiskFragmentation.pdf afs/ImpactofDiskFragmentation.pdf
bc173f6b4aa4408447f891bd0e3ed226  ImpactofDiskFragmentation.pdf
81f74b0da0958966da0afb8f0074ff26  afs/ImpactofDiskFragmentation.pdf

And on the server, I found that file under /vicepa:

# md5sum /vicepa/AFSIDat/4/4+++U/+/+/4+++++B
bc173f6b4aa4408447f891bd0e3ed226  /vicepa/AFSIDat/4/4+++U/+/+/4+++++B

So the file is clean when read from the server, and also from another IA32 
client.

If I repeat 'fs flush' on the file from the AMD64 client, I get a 
different checksum every time:

$ md5sum afs/ImpactofDiskFragmentation.pdf
e38228aea3fb1eb6af6677e6f8d154c1  afs/ImpactofDiskFragmentation.pdf
$ fs flush afs/ImpactofDiskFragmentation.pdf
$ md5sum afs/ImpactofDiskFragmentation.pdf
3f466fc07f1c9fa744733cec252c257f  afs/ImpactofDiskFragmentation.pdf
$ fs flush afs/ImpactofDiskFragmentation.pdf
$ md5sum afs/ImpactofDiskFragmentation.pdf
7ff29b7109e66f0ebb3876133c51e8df  afs/ImpactofDiskFragmentation.pdf

The server is an IA32 running Debian stable, with Debian kernel 2.4.27, 
and OpenAFS 1.4.0 binaries from backports.org.

The faulty client is an AMD64 (dual core) running Debian unstable, Debian 
kernel 2.6.15, and OpenAFS 1.4.1 binaries from amd64.debian.net.

The good client is an IA32 running Debian testing, vanilla 2.6.14 kernel, 
and OpenAFS 1.4.1 binaries from debian.org.

The kernel modules on each host were built with module-assistant.

I can help with debugging if needed (I do some kernel hacking for a 
living).


Cheers,

-- 
Jean-Marc Saffroy - jeanmarc point saffroy chez free point fr