Weekly edition	Kernel	Security	Distributions	Contact Us	Search
Archives	Calendar	Subscribe	Write for LWN	LWN.net FAQ	Sponsors

fall-through dentries

Posted Mar 26, 2009 16:10 UTC (Thu) by arnd (subscriber, #8866)
Parent article: Union file systems: Implementations, part I

AFAICT, the fall-through idea is really interesting but may get
problematic when the lower-level file system is modified. The original
union mount code should be able to handle readdir (after a new opendir)
and lookup even if the lower file system is bind-mounted to another
location and updated concurrently.

If you duplicate the entire directory structure in the top level of the
union, this would not even easily work in the case where you unmount the
top level, modify the lower level and then recreate the union mount.

(Log in to post comments)

Posted Mar 26, 2009 17:03 UTC (Thu) by vaurora (guest, #38407) [Link]

I think the key to getting a maintainable unioning file system is limiting the feature set. Okay, Linux union mounts won't slice, dice, AND puree your files - but they will cover many common cases.

If you don't agree, just wait for the aufs article - it is almost certain to implement any feature you want.

fall-through dentries

Posted Mar 27, 2009 14:36 UTC (Fri) by arnd (subscriber, #8866) [Link]

I absolutely agree that the feature set needs to be limited, and that's
what makes the Plan9 way of union mounts so beautiful (thanks for
describing it here, I didn't know how it works before).

Limiting the implementation so that you can never change the underlying
file system any more may be worth it but is still quite drastic, so I
thought it should be mentioned more explicitly.

fall-through dentries

Posted Mar 30, 2009 8:33 UTC (Mon) by bharata (subscriber, #7885) [Link]

> If you duplicate the entire directory structure in the top level of the
> union, this would not even easily work in the case where you unmount the
> top level, modify the lower level and then recreate the union mount.

Not only that but since you cache the consolidated directory entries on disk, you will not be able to union mount your top layer later on any other lower layer filesystem.

I know that supporting all sorts of corner cases and features has caused major pains for union mount, but this restriction sounds a bit too restrictive to me :)

fall-through dentries

Posted Mar 30, 2009 8:45 UTC (Mon) by bharata (subscriber, #7885) [Link]

> I know that supporting all sorts of corner cases and features has caused > major pains for union mount, but this restriction sounds a bit too
> restrictive to me :)

This also will cause problems if you want to use union mount in server consolidation environments where you have multiple servers working out of a common base distribution as their lower layer. With your scheme, I will not be able to do updates (like security updates) to the base distribution and see it getting effected in all the servers.