[Linux-cluster] takeover, fencing & failback

[Lon Hohberger]

On Tue, 2008-07-01 at 12:58 +0200, Gerhard Spiegl wrote:
> Hi all,
> 
> I'm working on a two node cluster (RHEL 5.2 + RHCS) with one
> XEN virtual machine per node:
> 
> node1 => VM1
> node2 => VM2
> 
> When node1 takes over VM2 via the command:
> 
> clusvcadm -M vm:VM2 -m node1
> 
> node2 gets fenced after takeover is done, which is probably expected behaviour.

No, it's not.

> As node2 comes up again it fetches his VM2 back (nofailback="0", but also
> fences node1 (ipmilan) where VM1 is still running an therefore interrupted and 
> restartet on node2.

Neither is this.  Fetching the VM back certainly should require
fencing...


> When node1 comes up the same game in the other direction begins.
> Is there a way to avoid this fence loop?


> In other words: can a service be migrated from node1 to node2 without other
> services that run on node1 being interrupted?

We'll need more details in order to figure out what's going on; such as
cluster.conf and your network topology (switch make/model, what speed
are your network links, etc)

-- Lon