[Linux-cluster] Fencing fibre switches

[Adam Manthei]

On Wed, Jun 08, 2005 at 04:24:37PM -0500, JACOB_LIBERMAN at Dell.com wrote:
> I've looked over the fence_* scripts for Fibre switches and I have a
> quick question. Dell sells McData and Brocade Fibre switches. Our
> customers always used name based zoning, rather than port based zoning.
> Therefore, csmtrs occasionally move their host initiators into various
> ports on the switch.
> 
> Are there any fencing mechanisms that fence based on the initiators
> World Wide Name rather than the port number?
> 
> Thanks, Jacob

No because when we wrote the agent, name based zoning was what brodace
called "soft zoning" and port based zoning was "hard zoning".  Soft zoning
meant that HBAs couldn't see the device when it logged into the switch, but
if it was writing to the disk, soft based zoning allowed it to continue.

The hard zoning made sure that all I/O would be stopped when the port was
removed from the zone.  

The fence_brocade script does not use either.  It actually just calls
portdisable to shut off the port.  The reason that this is used instead of
zoning is because the brocade firmware that we were using did not work real
well with dynamically changing zones while i/o was happening.  This
typically resulted in hung i/o which then hung the cluster.  The
disadvantage to this approach is that the port state, whether disabled or
enabled is not persistent across switch reboots.  This could potentially be
a problem if nodes are still up after being fenced and the switch gets
rebooted.  

Hope that helps.
-- 
Adam Manthei  <amanthei at redhat.com>