[OpenAFS-devel] fileserver problem

[Thomas Mueller]

On Fri, 14 Dec 2001, Kuba Ober wrote:

> > > I do not know if OpenAFS nowadays has protection against clients wi=
th
> > > one-way-connectivity using up all resources (threads or whatever) o=
f
> > > the server. The IBM versions don't seem to have such protection.
> >
> > I talked to the admin of the client and he told me that the machine
> > is running WinNT and had installed two different versions of McAfee's
> > VirusScan software. It was configured to scan all the network drives.
> >
> > So I think, there were two problems:
> > - a newer version of VirusScan was installed without a deinstallation=
 of
> > the old version
> > - the software was configured to scan all drives.
> >
> > Nevertheless should the server protect himself against a client
> > using up all ressources of the server.
> > I think there were a lot of scenarios where a client with an large AF=
S
> > cache could simply read all the files he could reach.
>=20
> Regardinbg one way connectivity: things can go mad, and that doesn't pe=
rtain=20
> only to AFS (NFS complains of not being able to allocate a slot, for ex=
ample,=20
> which doesn't mean much at first glance), when somewhere on your switch=
 or=20
> hub, there's a duplex disagree. We have a server connected to client ma=
chines=20
> via HP Procurve managed switch, and that switch, for whatever reason,=20
> sometimes switches the server's port to half-duplex mode. I had to disa=
ble=20
> autonegotiation and set kernel module options for the eepro100 card on =
the=20
> server to fixed full-duplex, 100mb/s, and set it same on the switch. Al=
l of=20
> our network cards are same eepro100 models. For some reason the switch =
would=20
> autonegotiate fine, and then after several hours, it would switch the p=
ort's=20
> mode to half duplex. As the network card was maintaining its fullduplex=
 mode,=20
> most of the packets were getting lost, sometimes with either upstream o=
r=20
> downstream part being affected more. It looked *very* strange and was a=
=20
> nightmare to debug. I was suspecting everything but our new switch...

Thanks for your comment, we have checked our switches and found no=20
problems here.

But we had the problem again during the last few days.

This time it was caused by the "Corel MEDIA FOLDERS Indexer"=20
- C:\Corel\Graphics8\Programs\MFindexer.exe - running from the Autostart=20
menu on a Windows NT 4.0 Workstation.

Again, the problem only appears with volumes holding a really large=20
number of files which are readable by system:anyuser on OpenAFS=20
fileservers (we have seen it on i386_linux22 und sun4x_58).

cbd shows that all of the CB structs are consumed by this client=20
and are related to different files within one volume.

Thomas.
--=20
--------------------------------------------------------
Thomas M=FCller, TU Chemnitz, URZ, 09107 Chemnitz, Germany
--------------------------------------------------------