Re: [Linux-cluster] Using cman,etc for a non-gfs app

[Lon Hohberger <lhh redhat com>]

On Wed, 2005-07-20 at 14:31 -0400, Olivier Crete wrote:
> On Wed, 2005-20-07 at 13:49 -0400, Lon Hohberger wrote:
> > On Fri, 2005-07-15 at 18:02 -0400, Olivier Crete wrote:
> > > On Wed, 2005-22-06 at 18:36 -0400, Lon Hohberger wrote:
> > > > * libmagma provides cp_fence() / clu_fence() which work on both CMAN and
> > > > gulm.
> > > 
> > > I've looked implementing clu_fence() in libmagma. It seems to only call
> > > an ioctl in the kernel cman which only seems to do cooperative fencing
> > > and removing a node from the cluster... I dont see any call to fenced or
> > > fence_node. Or have I missed something ?
> > > 
> > 
> > Correct, it removes the node via libgulm or asking CMAN to expire and
> > fence the node.
> 
> But cman does not seem to really fence the node... just remove it from
> its list..  So that's not enough for our app that has shared storage..
> Ideally, I would need the same behavior as stonith and have the call
> block and return only when we are 100% sure that the node have been
> fenced.

It's supposed to fence it.  At least, it used to fence it...  I wonder
if something changed that I didn't know about?

Basically, the idea is that we tell CMAN to force-expire the node.  It
should then take recovery actions to restore the node.

fenced is running, right?

-- Lon