[Linux-cluster] CS4 have a splitt-brain-problem in 2-node setup

[Falk Hackenberger - MediaTransfer AG Netresearch & Consulting]

Lon Hohberger schrieb:
> On Tue, 2006-05-23 at 12:21 -0400, Lon Hohberger wrote:
> 
>>On Mon, 2006-05-22 at 11:57 +0200, Falk Hackenberger - MediaTransfer AG
>>Netresearch & Consulting wrote:
>>
>>>Hello,
>>>
>>>I try to use the CS4 for managing a 2-node-cluster.
>>>also i have a working fenicing-device (apc-powerswitch).
>>>I think I have found a design-error in the CS4.
>>>
>>>If you boot up one node(1) and this node(1) cant reach the other node(2)
>>>node(1) trys to fence node(2).
>>>This is fine...
>>>but if the fencing fails, node(1)
>>>starts his services via rgmanager without to be shure that the other
>>>node is not working.
>>>volia and you have a splitt-brain problem.
>>
>>>any solution for this?
>>
>>rgmanager should not start services unless the node is quorate.  The
>>cluster should not ever become quorate if fencing fails.  Sounds like a
>>bug.

> Wait, fencing fails or fenced fails to start?
fenced start but fencing fails
simple way to reproduce it:

just put the node(1) in a other vlan, so that he can not reach his fence
device or the other node(2)
then reset node(1) do not reboot him...