Re: [Gnu-arch-users] extended attributes

[Chris Mason]

On Tue, 2004-01-06 at 20:25, Andrew Suffield wrote:
> On Tue, Jan 06, 2004 at 04:47:28PM -0500, Chris Mason wrote:
> > On Tue, 2004-01-06 at 16:42, Tom Lord wrote:
> > >     > From: Colin Walters <address@hidden>
> > > 
> > >     >> and a flat-namespace of attributes.
> > > 
> > >     > At least on Linux now, the namespace is subdivided at the toplevel; 
> > > it
> > >     > is separated into "user", "system", and "security".  The difference
> > >     > between "system" and "secure" is that the security namespace is only
> > >     > used for LSM modules.
> > > 
> > > It's primarily "user" and secondarilly "security" (i thought they were
> > > called "trusted") attributes that I object to.   "System" attributes
> > > seem actually useful (ACLs being an example).
> > 
> > People have been debating EA usage and semantics for quite a while, and
> > that's not going to change any time soon.  For right now, I'd suggest
> > something like this:
> > 
> > 1) support the acls, they're well understood
> 
> The right way to support ACLs on linux is to support EAs. Otherwise
> you either waste time implementing the same stuff N times over, or you
> only support one ACL scheme.

I agree EAs are the right way to support acls in general, which is why
we went that way for the reiserfs code.

The real question once you talk about supporting them in arch is how do
you store them in the changeset.  ACLs are easy because they are just
like a change in permission bits (replacement without diffing).  Generic
eas are harder because you can have text files with binary eas etc, so a
simple diff won't work.

If diff and patch support the EAs, you get 90% of the work done for
free.  Functions to copy files around still need tweaking, but that
should be minor.

-chris