Subject: Re: Thoughts on getting rid of obj links
To: None <current-users@NetBSD.ORG>
From: der Mouse <mouse@Collatz.McRCIM.McGill.EDU>
List: current-users
Date: 12/20/1995 06:54:16
>> That's all well and good, but you can't union mount over NFS, which
>> apparently some people must do if they can't afford local disk space
> You can - if union mounts were changed to not require whiteouts.
> You would not be able to remove a file that existed in the lower
> layer - but that would be fine for this situation.

Actually, there is no _requirement_ for whiteouts at all.  It's just
the way the current implementation works.  There's no need to break
UNIX filesystem semantics even worse than unionfs already breaks them.

I did a user-level NFS server implementation of a real union
filesystem, with proper semantics for everything regardless of where
files live (even better than NetBSD union fs).  It kept its metadata in
upper-layer files, and hid those files from its clients.  It worked as
well as anything could over NFS.  (I started using it for /usr/src and
rapidly found out that a whole lot of stuff simply didn't work right;
investigation forced me to blame NFS rather than my code.)

It could probably be turned into a kernel implementation, but without a
working filesystem to use as a model, I didn't feel up to taking on
that task.  (nullfs didn't do locking right; perhaps I could have used
the existing unionfs as a model....)

					der Mouse

			    mouse@collatz.mcrcim.mcgill.edu