Btrfs devel - Oct 2007 - xattr design thoughts

Hello,

Now that I actually understand whats going on with the generic xattr handling
stuff, I've rethought how to do xattrs.  Currently each individual xattr for
a
particular inode has its own item, or if its too large its own inode.  So if I
create file blah and then add "security.mysec",
"user.stuff", "trusted.whatever"
to blah, there will be 3 dir_item's, one for each xattr.  I was trying to
figure
out why ext3 wasn't storing the prefix on disk with the xattr and I figured
it
out, it has its own little object for holding xattrs for the inode.  So instead
of having 3 different locations for its xattrs, it has one big chunk that it
uses for its xattrs.  So this is what I was thinking of doing, having a xattr
disk structure that holds the name, the index (for looking up the
prefix/handler), and the value.  Then we only use one dir_item per inode for
xattrs, and just pack in new entries as they come along, and then once we
outgrow the dir_item just move it all into an inode.  Let me know what you guys
think about this.  Thank you,

Josef

On Tue, 2007-10-16 at 10:26 -0400, Josef Bacik wrote:
> Hello,
> 
> Now that I actually understand whats going on with the generic xattr
handling
> stuff, I've rethought how to do xattrs.  Currently each individual
xattr for a
> particular inode has its own item, or if its too large its own inode.  So
if I
> create file blah and then add "security.mysec",
"user.stuff", "trusted.whatever"
> to blah, there will be 3 dir_item's, one for each xattr.  I was trying
to figure
> out why ext3 wasn't storing the prefix on disk with the xattr and I
figured it
> out, it has its own little object for holding xattrs for the inode.  So
instead
> of having 3 different locations for its xattrs, it has one big chunk that
it
> uses for its xattrs.  So this is what I was thinking of doing, having a
xattr
> disk structure that holds the name, the index (for looking up the
> prefix/handler), and the value.  Then we only use one dir_item per inode
for
> xattrs, and just pack in new entries as they come along, and then once we
> outgrow the dir_item just move it all into an inode.  Let me know what you
guys
> think about this.  Thank you,
For the common case where we have a really small number of xattrs on an
inode, I doubt it makes a huge difference either way.  The dir items
should be cheap enough that you can create one dir item per xattr
without much pain, and I prefer that approach over one blob.

The problem with the blob is that the performance of the xattrs as a
whole is dependent on your largest xattr.  With the one dir item per
xattr, the btree takes care of all of that nonsense for you.

-chris