[OpenAFS-devel] Re: Integrating OpenAFS into Linux (and RedHat)

[Nathan Neulinger]

Derek Atkins wrote:
> 
> Nathan Neulinger <nneul@umr.edu> writes:
> 
> > Ah... Yeah, coda is different. Coda is implemented such that the kernel
> > module is just an relay to the user space coda/venus daemon. I believe
> > xfs in Arla is the same way. They generalized the filesystem interface,
> > and then a user space task simply attaches to the kernel device node to
> > provide the actual filesystem support.
> 
> Ahh, well, I suppose we could do something like that, then...  I'm
> not sure *I* want to do that, tho :)

It does obviously mean a performance hit. But - it would certainly be a
useful option, i.e. allow a linux afsd option that would simply attach
to an XFS or CODA device instead of loading a kernel module. That way,
you get the benefit of having AFS be universally available without a
kernel mod, and anyone else who wants a performance boost could just go
and load the kernel module directly.

The XFS approach by arla does have one definite benefit - it makes
porting to new architectures extremely easy. All you have to do is port
the xfs module. Granted, you have something similar in porting libafs.o,
but it's alot easier to port and maintain a simple small module that it
is to maintain a large one like libafs. One other benefit of the arla
approach is that you can even run arlad WITHOUT connecting to the kernel
- so it acts like an ftp client for testing purposes. If I remember
correctly, the arla team has the pure userspace portion working on
almost any autoconfable system, it's just the kernel portion that needs
ported.

-- Nathan

------------------------------------------------------------
Nathan Neulinger                       EMail:  nneul@umr.edu
University of Missouri - Rolla         Phone: (573) 341-4841
CIS - Systems Programming                Fax: (573) 341-4216