Usefulness of extended attributes over NFS

[Alexander Larsson]

On Mon, 2010-06-28 at 16:07 +0200, Lennart Poettering wrote:
> On Mon, 28.06.10 09:18, James Morris (jmorris at namei.org) wrote:
> However, that's mostly where he story ends I think, it is indeed not
> widely used. Generally I believe they are useful however, and if they
> would be ubiquitiously available they'd probably be used more
> often. However, for that to happen we'd also need something like a
> fpathconf() check or so to figure out whether user xattrs are allowed or
> not.

I don't think thats quite the full story. Even when they are supported
on a filesystem xattrs have problems that make them non-ideal in
practice for large-scale desktop use.

For small attributes like selinux labels the default ext3/4 inode size
is large enough that the xattrs fit in the inode. But as soon as you add
any more data to it, the xattrs won't fit in the inode. In my
experiments as little as 4 bytes + a selinux label made it go outside
the inode.

This means that to get this data you need an extra seek, and one extra
seek per file when doing a readdir operation is extremely costly.

And this performance problem is real, not some theoretical idea. I had a
discussion about this with Eric Sandeen in my gvfs metadata blog post
(http://blogs.gnome.org/alexl/2009/06/24/data-about-data/ - see the
comments near the end). 

There a simple test breaks down completely, performance-wise when using
xattrs. Reading a directory with 10000 entries with mime sniffing took 6
seconds, adding xattrs made it take 40 seconds.

-- 
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
 Alexander Larsson                                            Red Hat, Inc 
       alexl at redhat.com            alexander.larsson at gmail.com 
He's a sword-wielding dishevelled photographer who dotes on his loving old ma. 
She's a cold-hearted green-skinned snake charmer with an evil twin sister. 
They fight crime!