[Linux-cluster] RE: < fecing with out any hardware? >

[David Teigland]

On Sat, May 06, 2006 at 10:05:35PM -0500, Stanley, Jon wrote:
> The question that I have is that there is functionality in the SCSI-3
> spec for Persistent Group Reservations.  Basically, what happens is that
> each system that wants access to a disk puts a "reservation" and
> "registration" on it.  A commercial clustering solution (Symantec) uses
> this feature in order to do it's I/O fencing.
> 
> The initial reservation on the disk is "Write Exclusive Registrants
> Only", meaning that if you are not registered to be on the disk, you
> cannot write to it.  When the node comes up, upon synchronizing with all
> of the other nodes, etc, it puts it's key onto the disk.  It can then
> write to the disk, without any problem.  When the node dies, the
> surviving node(s) see that, and eject the dead node, making it
> physically impossible to write to the disk.
> 
> This of course requires support from the array to do it (it's a SCSI-3
> standard, but not all arrays implement it), thereby limiting the choice
> of storage to mid-to-high-end enterprise arrays.  
> 
> The question is why can't we use that as a fence mechanism, and do away
> with the hardware poweroff stuff, if the array supports it?  Of course
> the hardware poweroff stuff could be left in for older/lower end arrays,
> etc, but I think that options are a Good Thing(TM).

You could definately use persistent reservations to do fencing, we just
don't have a fencing agent written for it yet.  It's one of those things
that no one ever quite gets the time to do.  It's something that would be
_really_ nice to have and would spare a lot of people a lot of hassle.

Dave