Jochen,

On Monday 18 May 2009, Jochen Roderburg wrote:
> Zitat von Jochen Roderburg <[email protected]>:
> > Situation on two systems (one at home, one at work):
> >
> > VirtualBox 2.2.0
> > Linux Host (Slackware-based with self-compiled stock kernel)
> > Windows 2000 Guest
> > Bridged networking
> >
> > Network Chips:
> > on home system: Realtek Semiconductor Co., Ltd. RTL8111/8168B PCI
> > Express Gigabit Ethernet controller
> > on work system: Intel Corporation 82566DM-2 Gigabit Network Connection
> >
> > Everything works fine on both systems with Linux kernel 2.6.28 and
> > also on the home system with 2.6.29.
> >
> > But with a 2.6.29 kernel on the work system the direct network
> > connections between guest and host are extremely slow and not usable.
> > Examples:
> > A longer output in a terminal window comes with great pauses and
> > eventually stalls.
> > Filesystem connections to a samba server can't display large
> > directory listings, in the samba log files I see timeout errors and
> > frequent server restarts.
> >
> > Astonishingly at the same time connections from the guest to other
> > remote systems work without problems.
>
> Seems I have to answer this myself.  ;-)
>
> In short: it was a bug in the 2.6.29 kernel which is already repaired
> in kernel version 2.6.29.2.
>
> More explanations:
>
> I wrote this question both in the vbox forum and here, but did not get
> an answer in neither places. But in the forum some other people
> reported problems which very much looked like the same and one person
> had found the essential clue, namely that it had something to do with
> the gso (generic segmentation offload) feature of the Linux kernel.
> And with this information it was not difficult to google a redhat
> bugzilla entry https://bugzilla.redhat.com/show_bug.cgi?id=490266
> where the whole problem was discussed, explained and solved.
>
> It was/is a bug in the linux kernel 2.6.29 which was triggered by the
> tun driver (or perhaps the vbox netfilter in our cases) in connection
> with nic drivers that use that gso feature.
>
> The (one-line) patch is now in the kernel sources since 2.6.29.2.
>
> So we have various solutions now:
> 1) A workaround is to turn off the "segmentation offload" feature for
> the nic (ethtool -K ethn tso off seems to do the trick).
> 2) Add the one-line patch to your kernel sources yourself.
> 3) Use a kernel version from 2.6.29.2 up.
>
> I have tested #2 now on my system and the network connections between
> host and guest work fine again.

Thank you _very_ much for this information! Actually we saw several
reports about this issue and we were able to reproduce it locally
as well. Disabling the tso/gso feature with ethtool helped here but
this isn't applicable of course. So far we assumed that this was due
to some bug in our code but I was able reproduce the issue locally
with 2.6.29.1 and can confirm that 2.6.29.3 restored the full speed.

Thanks again,

Frank
-- 
Dr.-Ing. Frank Mehnert    Sun Microsystems, Inc.    www.sun.com

Attachment: signature.asc
Description: This is a digitally signed message part.

_______________________________________________
vbox-users mailing list
[email protected]
http://vbox.innotek.de/mailman/listinfo/vbox-users

Reply via email to