[Pkg-xen-devel] Bug#679533: Bug#679533: Bug#679533: Traffic forwarding issue between Xen domU/dom0

[Thomas Goirand]

On 12/08/2012 07:32 AM, Hans van Kranenburg wrote:
> So...
> 
> After disabling hyperthreading in the server bios we can no longer
> reproduce the bug. This is a HP Proliant DL360 G7. We've not yet tried
> to consistently reproduce on other hardware we have, but all production
> hosting servers we run now are of these G7 type with Intel Xeon X5650 or
> X5675 cpu.
> 
> The last test Frank reported, starting >20 of vms and in a loop
> rebooting them consistently reproduced the broken behaviour on a random
> network interface of a random vm.
> 
> We're now at >150 reboot cycles of all test vm's on a single dom0 with
> HT disabled. When enabling HT this fails consistently between 10-50 vm
> reboots.
> 
> So this still looks like some race condition bug. It's reproducible
> while running with either openvswitch or linux bridge, so it's only
> related to the actual dom0/domU passthrough of network traffic.
> 
> We're still gonna try to upgrade the test dom0 to wheezy with xen 4.1
> and linux 3.2 and see what happens with HT enabled and disabled.
> 
> I'm very interested in feedback from Xen/Linux developers about this.
> Enabling HT with Xen 4.0 and Linux 2.6.32 on debian is not safe right
> now. Been there, done that. :-)
> 
> During the past months, it seemed we're the only ones affected by this
> bug. Or, no-one is starting, stopping and live migrating xen vms as much
> as we do?
> 
> Thanks,
> Hans

Hi Hans,

I thought I should post to this bug to let you know you're not alone.

I have seen the same kind of networking flaws you talked about, with
loss of connectivity. I believe this is a very old bug in Xen, and that
even Xen 3.x is affected. The thing I've done to fix it was to run ping
requests of all VMs from dom0 on a cron job. It also seems that ipv6 is
affected as well.

What you've done to be able to reproduce the bug is very nice, and I'm
sure it will somehow help finding where the problem lies. Your finding
about HT is also very nice.

Cheers,

Thomas