On Tue, Feb 20, 2001 at 08:24:27AM -0800, [EMAIL PROTECTED] wrote:
> [NOTE:
> The full quote of the original message has been removed.
> --Andreas Gruenbacher.]
>
> Hi Andreas,
>
> Your proposal would certainly do the job for SnapFS and for a variety of
> things we are planning with InterMezzo (global fid's, versioning).
>
> Here are a few comments:
>
> There is little discussion of "out of EA space" conditions. Is there a
> maximum size? If there are ACL and snapshot EA in the same file system, how
> do we agree on what is "max" between the two EA's?
>
> Should you get to define inheritance, then it would be important to note
> that we wouldn't really want to inherit any attributes from other inodes: if
> we do, we would likely wipe them clean immediately.
>
> I am a strong fan of your interfaces for atomicity. While it complicates
> the interfaces and while the definitions of atomicity may need some
> refinement as Robert notes, not having atomicity can lead to an interface
> that is insufficiently powerful.
>
> I agree with many of Robert's other observations. It does seem to me that
> there is value in separating out the interface for ACL's and EA's and also
> that POSIX ACL's do not deserve special treatment. I also think that a
> "writev" style interface will be much easier to use, and the reading of
> attributes does need to be defined.
>
> I'd also be pretty happy with the "Proposal" made in Andrew Gilfind's paper,
> although again I think that the
> emphasis of POSIX ACL's over others is not right.
>
This isn't true, an absolute requirement for the interface I proposed
was that it be extensible in the set of ACLs that it supports.
Another aim, however, was that a filesystem that does provide extended
attributes can implement support for a range of ACL types almost for free
by drawing on a common infrastructure for mapping ACLs to particular
extended attributes and on-disk formats.
In the proposal I identified 8 common steps ACLs undergo on their
way into the filesystem. On the way in: validation, updating in-memory
caches, conversion to on-disk format, storage into the filesystem.
On the way out: retrieving in-memory cached data, retrieving from
the filesystem, conversion back to in-memory form. And finally
enforcement. Each of these steps is then supported by common code that
provides default functionality *for a range of supported ACL types*
including POSIX ACLs. The purpose of this common infrastructure is to
make the common cases easy, however, any filesystem can choose to ignore
some or all of this common functionality as it sees fit.
Ideally for XFS we basically want a vanilla ACL infrastructure sitting
above us that works out how to map incoming ACLs of different classes
to extended attributes and then simply asks us to update those attributes.
In most cases we can simply accept what that common infrastructure
gives us, but for POSIX ACLs we need to be able to control the on-disk
format of the data (for interoperability with IRIX XFS filesystems).
Ideally we should be able to override that function, but let other,
common code, handle all of the other cases.
What I was trying to achieve with the proposal in the document was
to get as much ACL support as possible (for as large a range of ACL
types as possible) without dragging large chunks of ACL support code
into each and every filesystem.
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]