[Linux-cluster] rgmanager "depend" tag on RHEL4

[Lon Hohberger]

On Wed, Jul 25, 2007 at 08:50:17PM +0200, Mark Hlawatschek wrote:
> > > I'd like to use the "depend" feature of rgmanager on RHEL4. Is there a
> > > way to do this. E.g. would it be possible to run the cvs HEAD version of
> > > rgmanager on a RHEL4.5 cluster infrastructure ?
> >
> > It doesn't build on RHEL4 right now; there's no equivalent to
> > 'cman_get_fenceinfo()'.
> >
> > If you #if 0 out the 'node_fenced()' body and return 1, it builds.  I
> > don't know if it will run... but it builds ;)
> >
> > It doesn't appear there's an equivalent function; so to implement it,
> > one would have to perhaps parse /proc/cluster/services or something like
> > that...
> What does the prototype of cman_get_fenceinfo() look like ? What is it meant 
> to do ?

It gets the time the node was last fenced from CMAN.  When a node
rejoins the cluster, this information is cleared.

So, what rgmanager does is effectively:
  nodeX died
  does nodeX have fencing configured?
    yes
      wait for node X to be fenced or rejoin the cluster
    no
      continue

> Is there another way to get the "depend" feature back-ported ?

I think there's already a feature request open about it, by you?

Implementation-wise, in RHEL5, there's an event queue for service state
changes, and a function which broadcasts starts/stops
(broadcast_event()).  The difference is that in RHEL5, you can multicast
to all rgmanagers w/ one call - whereas on RHEL4, you have to open a
connection to each and send the message individually.

So, basically, it involves porting the event queue code and implementing
broadcast_event().

-- 
Lon Hohberger - Software Engineer - Red Hat, Inc.