[OpenAFS] Problem with cell name aliases in CellServDB

[Jeffrey Hutzelman]

On 27 Aug 2001, Derek Atkins wrote:

> Honestly, I never knew that "cell name aliases" was ever supported.
> I'm not at all surprised that it fails.  Think about it:  what does
> a cell alias _mean_?  Are you talking to btv.ibm.com or micro.ibm.com?
> Which token do you get?

Come on, Derek.  A feature is not automatically broken just because you
don't understand what it does.  Still, I'll explain...

The correct name for the feature we're talking about is "linked cells".  A
CellServDB entry with two cell names like the one Markus quoted actually
describes only the first cell named, but it also designates the two cells
as "linked cells".  This relationship is symmetric and cannot be applied
to a group of more than two cells, though it is possible to have multiple
non-overlapping pairs.  AFAIK this has never been a supported feature, and
I had no idea anyone was using it.  It has two effects on the way the
cache manager processes mount points: 

- If a VLDB lookup in one linked cell fails, the operation will be retried
  in the other cell.  So, effectively, a mount point in one of these cells
  can refer to a volume in either cell.  This process happens only at the
  time a mount point is evaluated, so it can't be used for some kind of
  cross-cell failover; it's only useful when you don't know which cell a
  volume will be in (for example, during a transition).

- When deciding whether to revert to an RO path, the cache manager treats
  volumes in the linked cells as being in the "same" cell.  So, crossing
  a cellular mount point from one linked cell to the other does not
  automatically result in reverting to a read-only path as it would in
  other cases.

-- Jeffrey T. Hutzelman (N3NHS) <jhutz+@cmu.edu>
   Sr. Research Systems Programmer
   School of Computer Science - Research Computing Facility
   Carnegie Mellon University - Pittsburgh, PA