[Linux-cluster] scsi reservation issue with GFS

[Ryan O'Hara]

On Thu, Feb 19, 2009 at 04:00:21PM -0600, Alan A wrote:

> Thank you very much, that did the trick. I will remove sg3_utils package
> now. Hopefully this will never happen again. I had to remove SCSI
> reservations from 4 different volumes. Any suggestions on how to avoid this
> in the future?

I'm not sure why you experienced this problem. That said, there was a
problem in RHEL4 where scsi_reserve was enabled by default. The
fallout is/was that reservations would be created (on all devices that
supports them) whether you wanted them or not. Since then,
scsi_reserve has been modified to look at your cluster.conf, and if it
does not find fence_scsi that the fence agent, it stops before is
creates any reservations. This check is there solely to prevent
unintended SCSI reservations.

> We were planning on using SCSI fencing in combination with
> APC Power Switches, I am not sure this is such a good idea after this
> experience? Any additional thoughts?

I don't see why it wouldn't work.

Keep in mind that all of your devices must exist on cluster volumes
and on arrays that support SCSI reservations. There is a tool that
helps test this called 'fence_scsi_test'. I don't recommend that you
run the test script on a production cluster, though, since it will
attempt to create and then remove SCSI reservations on all
volumes.

Regarding the problem that you had with reservation conflicts ... it
appears that the nodes reporting the errors was no registered with the
device(s). So what you were seeing is expected behavior.

There is a document I wrote that explains how this all works. Here is
a link.

http://sources.redhat.com/cluster/wiki/SCSI_FencingConfig