Re: recursive spinlocks. Shoot.

[Carl-Daniel Hailfinger <c-d.hailfinger.kernel.2003@gmx.net>]

Robert White wrote:
> 
> Here is a problem statement that is virtually unimplementable in the current
> Linux kernel because of the simple-minded locking model.
> 
> Create a meta file-system that takes two existing (backing) file system
> devices (or existing mounted directories) and aggregates them such that:
> 
> (Ladies, and Gentlemen, The "deltafs" file system...)

Call it unionfs.

> 1) It doesn't matter what types of file systems are used as the backing file
> systems.
> 2) The aggregate file system is fully read-write
> 3) The base (first existing) file system is read-only
> 4) The front (second existing) file system is read-write
> 5) All operations are available on the aggregate file system (unlink,
> rename, open for write, open for append, chown, set access time, etc.)
> 6) The aggregate file system engine will transcribe all the modified files
> from the base to the front file system if the file is modified.
> 7) The aggregate file system will (probably using a reserved file name and a
> journaling structure encoded therein) maintain a white-out list to hide
> unlinked and renamed files.
> 8) After unmount, the front file system should be a minimal delta of the
> base file system
> 9) Remounting the same combination of file systems should consistently
> result in the same, consistent file system image.
> 10) (you get the point... 8-)

IIRC, Al Viro was working on this and we might have it in 2.6

http://www.ussg.iu.edu/hypermail/linux/kernel/0201.0/0745.html

Al?


Regards,
Carl-Daniel