On Mon, Sep 19, 2011 at 06:21:46PM +0200, RaSca wrote:
> Il giorno Lun 19 Set 2011 17:55:58 CEST, Dejan Muhamedagic ha scritto:
> > Hi,
> > On Wed, Sep 14, 2011 at 09:43:43AM +0200, RaSca wrote:
> >> Hi all,
> >> I've got a two node pacemaker/corosync cluster with some virtual domain
> >> resources on some DRBD devices.
> >> Every DRBD device is configured in dual primary setup and I have enabled
> >> the live migration. Cluster has also stonith enabled.
> >> My problem is that if a live migration for a single virtualdomain
> >> resource fails, then this node gets fenced, making unavailable also all
> > AFAIK, failing migration shouldn't result in node fence. I guess
> > that actually the subsequent stop operation failed, right? In
> > that case, that's probably a bug somewhere in the RA or VM code.
> > Thanks,
> > Dejan
> 
> Hi Dejan,
> thanks as usual for your response. In the end, since that I was facing 
> too much unexplainable problems I decided to upgrade libvirt and the 
> kernel itself to a newer version (from squeeze to squeeze-backports). 
> Until now problems seems to be resolved.
> 
> In Pacemaker Explained (Andrew, I'm almost finished with the 
> translation, I swear!) it is written that the default action on fail is 
> "fence", so it is assumed that if a single resource fails, then the 
> entire node is fenced. Note that at the moment every of my virtualdomain 
> resource have got the on-fail action set with "restart", and I've not 
> faced any fence.
> But please, help me to understand this: what do you mean with 
> "subsequent stop operation"?

Cluster may issue a stop operation if migrate_x fails.

> It is very plausible that this was the 
> reason, since the failed virtual machines were in state "paused" even if 
> I was forcing the stop. Does this is enough to make a node fence? Why 
> this failure is not considered in "on-fail" parameter declaration?

on-fail is a per-operation attribute. By default, it is set to
fence only for the stop operation. The point is that a failed
stop means that the cluster cannot establish the state of the
resource anymore, so the only remedy remaining is to fence the
node.

Thanks,

Dejan

> Do I made myself clear?
> 
> Thanks a lot,
> 
> -- 
> RaSca
> Mia Mamma Usa Linux: Niente รจ impossibile da capire, se lo spieghi bene!
> [email protected]
> http://www.miamammausalinux.org
> 
> _______________________________________________
> Linux-HA mailing list
> [email protected]
> http://lists.linux-ha.org/mailman/listinfo/linux-ha
> See also: http://linux-ha.org/ReportingProblems
_______________________________________________
Linux-HA mailing list
[email protected]
http://lists.linux-ha.org/mailman/listinfo/linux-ha
See also: http://linux-ha.org/ReportingProblems

Reply via email to