I don't know yet much about the implementation details of unionfs, but it 
sounds to a me as a general problem which arises if one tries to stack 
union-fs together. The problem is the fixed prefix in the name-schema for the 
whiteouts. If the prefix would be .wh.<fsid>. with <fsid> some sort of 
identifier which must be unique in the stack of filesystems, then the problem 
could be solved. Whiteout-names with the prefix .wh.1234. would be rejected 
of the unionfs-layer with the fsid=1234 returning EPERM, whereas a whiteout 
from the layer above with e.g. the fsid=5678 could be written.

Could this be a solution?

This would correspond to an option fsid=nnnn similar to NFS.


Am Mittwoch, 14. Dezember 2005 17:44 schrieb [EMAIL PROTECTED]:
> > O.k., I see the point. But: why does unionfs lookup a whiteout in a
> > read-only branch?
>
> While I was thinking about Wilhelm's question, I've got an idea of new
> option "dirs=<branch>=ro_wh".
> Currently these options are 'ro' and 'rw'. A 'ro_wh' option is
> perfectly equivalent to current 'ro'. But the meaning of 'ro' option
> changes. In case of 'ro', unionfs does not lookup a whiteout entry, but
> 'ro_wh' does.
> Reducing the number of lookup will get a chance to make unionfs
> lighter. I hope it will be effective to the simple cases, like tmpfs +
> large cd/dvd image, tmpfs + readonly nfs, tmpfs + regular mntpnt and so
> on. These cases will not need to modify the unionfs option. And I guess
> that most users would use unionfs like this.
> But it may not be effective in the complex cases, like the
> branch is specified readonly currently but it had been writable, or it
> is a writable branch of another unionfs. In these cases, user will need
> to change the option from 'ro' to 'ro_wh', to keep on looking-up on the
> readonly branch.
>
> I have once thought it doesn't need to lookup on the lowest branch only.
> But now I think most readonly branch does not need to lookup whiteout,
> and it is better to introduce a new option and make it choosable by
> user.
>
> How do you think about changing the meaning of 'ro'? > all of users and
> developers
> If I write a patch, will it be merged?
>
>
> Junjiro Okajima
> _______________________________________________
> unionfs mailing list
> [email protected]
> http://www.fsl.cs.sunysb.edu/mailman/listinfo/unionfs

-- 
--
Wilhelm Meier
email: [EMAIL PROTECTED]
_______________________________________________
unionfs mailing list
[email protected]
http://www.fsl.cs.sunysb.edu/mailman/listinfo/unionfs

Reply via email to