On Mon, Feb 19, 2001 at 09:50:14AM +0000, Stephen Tweedie wrote:
> Hi,
>
> On Sat, Feb 17, 2001 at 04:42:32PM +0100, Andreas Gruenbacher wrote:
> > Hello everybody,
> >
> > this is my third push towards evolving the extended attributes system call
> > interface to something that is acceptable for inclusion into the standard
> > kernel :-) I've tried to write things up as a manual page. The source and
> > a pdf version can be found at <http://acl.bestbits.at/pre/>.
> >
> > A mostly finished implementation for ext2 of the EA interface documented
> > here, and of ACLs on top of it, is almost finished. I'm planning to solve
> > the remaining quircks this weekend, so perhaps I can post a link tomorrow.
>
> This still seems to offer no way of solving the big problem I've been
> trying to stress for months now: you have namespace problems *inside*
> an ACL namespace.
>
> As a user using your POSIX acls, how can I set an ACL for some remote
> user coming in over NFS from a different Kerberos realm? It might
> even be a user using NFSv2: we don't have to consider only NFSv4 ACLs
> here. And we already have code to map remote kerberos tokens to
> local, temporary "process authentication groups": that's how AFS
> works.
>
> In other words, local ACL granting might be done in terms of these
> temporary unix gids, but (a) the permanent on-disk encoding of the
> username and (b) the username passed in over the ACL API both need to
> be expressed in terms of user@realm names.
>
> Not only NFS is affected by this: you might well allow remote kerberos
> users to login to the local node using an anonymous account, with the
> login process creating an appropriate authentication group for the
> remote user. You still want that user to be able to access files on
> the local filesystem with the appropriate ACL enforcement.
>
> Cheers,
> Stephen
>
Stephen if you haven't done so I'd really recommend you read the
document I wrote, one of the key things I worked on was understanding
where you're coming from and trying to connect you're approach
with the general approach advocated by Andreas and others.
The issues are probably better articulated there than I can manage in
a quick email..
There are two ways to view the way in which ACLs should be represented.
For want of a better way of describing them, I'll refer to them as
ACL-centric (the POSIX way) and ACE-centric (the NFSv4/NT way).
For the ACL-centric approach the ACL is the atomic object, it is
contructed in userland, passed into the kernel as a single blob,
and typically stored within a single extended attribute.
The ACE-centric approach views individual ACEs as the atomic
objects. There is a richer set of ACE types (alarm and audit),
and the ordering of ACEs affects the permissions granted for
the file. From my understanding of what you've posted and
you're proposed system call interface, it appears that you
expect ACEs to be stored one per extended attribute within
the filesystem.
What I would argue is that as long as you support and
store multiple classes of ACL, the ACL-centric approach
is equivalent to the ACE-centric approach,
and in fact I believe it offers some advantages.
In your proposal, in order to access and interpret a
collection of ACEs you provide the concepts of name family
and attribute class. The name family identifies the type of
the SID, and provides the mechanism through which you support
new identification types (blah@realm, etc.). However, to actually
do a permission check you need to interpret the "super-ACL"
that you've built with a particular semantics. For this you
provide the concept of an attribute family. Thus if I need to do a
get or set with POSIX semantics I pass ATR_POSIX_ACL.
So where are the equivalencies?
ACL-centric ACE-centric
----------- -----------
A POSIX ACL Interpret ACEs as a POSIX ACL
A NTv4 ACL Interpet ACEs as an NTv4 ACL
Update with POSIX Interpret ACEs as a POSIX ACL,
semantics: map new update, and then resolve any
POSIX acl to NTv4 ACL conflicts with existing ACEs
and store
New classes of ACL: Add new SID types and update the
add new ACL types the rules used to enforce
and provide support existing ACL semantics.
for converting to/from
existing ACL typs
The disadvantage of the ACE-centric approach is that for
filesystems such as XFS, that already store ACLs one
per-extended attribute, an ACE-oriented access method
will mean those filesystems must now crack open what
were once opaque objects (extended attributes) and
must now have some knowledge of the semantics of the
contents. This isn't a killer, but it violates the
idea that extended attributes should really just be blobs
of opaque data the filesystem keeps around for higher-level services
to interpret and make use of.
Again, I'd emphasis that the paper is probably more coherent
than this email, and I would really appreciate feedback you
have on the sections reviewing your proposals.
Andrew
--
Andrew Gildfind - R&D Software Engineer - SGI Melbourne Australia
email: [EMAIL PROTECTED] - work: +61.3.9834.8200 mobile: 0412.834.183
-
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to [EMAIL PROTECTED]