Re: [9fans] user-level file systems for Linux

Re: [9fans] user-level file systems for Linux

[YAMANASHI Takeshi]

On Thu Feb 19 11:01:55 JST 2004, Russ Cox wrote:
> There is a user-level file system driver for Linux
> called FUSE that has just released a new stable version:
> http://sourceforge.net/forum/forum.php?forum_id=350517
> They have user-level servers to mount various archives
> and various network protocols (http, ftp, smb).

I have been pondering about making a central file server
that serves a file system to PCs.  PCs are scattered around
the university and running Linux or Windows, maintained by
normal people; that is, PCs are really Personal Computers.

I'd love to see the server running Plan 9, but the protocols
between the server and PCs were an issue because NFS must
be avoided in this case.  But now, FUSE/ftp would settle
the issue for Linux.  Windows can mount ftp servers as if
they are normal folders already.

But still, I wish FUSE will support 9p2000 someday. :)
--

[9fans] user-level file systems for Linux

[Russ Cox]

There is a user-level file system driver for Linux
called FUSE that has just released a new stable version:
http://sourceforge.net/forum/forum.php?forum_id=350517
They have user-level servers to mount various archives
and various network protocols (http, ftp, smb).

The protocol itself looks fairly similar to the old 9P:
http://tinyurl.com/28f2y

There is another similar project called LUFS
http://lufs.sourceforge.net/lufs/
though it seems more complicated than necessary
and makes at least one (imo) critical mistake: no fids.

Actually there have been a handful of projects doing
similar things over the past few years, and they've all
died out.  I'm optimistic about FUSE because it's almost
an exact translation of the VFS layer, meaning that it's
simple and as expressive as possible.

Russ

Re: [9fans] user-level file systems for Linux

[boyd, rounin]

> I'm optimistic about FUSE because it's almost
> an exact translation of the VFS layer, meaning that it's
> simple and as expressive as possible.

hmm, interesting.  i'd like to code some f/s interface to some
PCI FPGA boards, which would make life a lot easier.

Re: [9fans] user-level file systems for Linux

[ron minnich]

On Wed, 18 Feb 2004, Russ Cox wrote:

> Actually there have been a handful of projects doing similar things over
> the past few years, and they've all died out.  I'm optimistic about FUSE
> because it's almost an exact translation of the VFS layer, meaning that
> it's simple and as expressive as possible.

my biggest problem when I did v9fs in 1996/7 was convincing anyone in
linux-land that user-level mounts were a good idea. I got the weirdest
arguments against both user-level mounts and private name spaces. It seems
like they've finally started to get it.

ron

Re: [9fans] user-level file systems for Linux

[boyd, rounin]

> my biggest problem when I did v9fs in 1996/7 was convincing anyone in
> linux-land that user-level mounts were a good idea.

ultrix would let you do it back in the early 90's, but it
was luser mode NFS -- yuk!

Re: [9fans] user-level file systems for Linux

[Martin C.Atkins]

On Wed, 18 Feb 2004 20:57:42 -0500 "Russ Cox" <rsc@plan9.bell-labs.com> wrote:
>...
> Actually there have been a handful of projects doing
> similar things over the past few years, and they've all
> died out.  I'm optimistic about FUSE because it's almost
> an exact translation of the VFS layer, meaning that it's
> simple and as expressive as possible.
>
> Russ

I wrote a user-mode filesystem for Linux just yesterday. In Python
(using a library I hacked up early last year). Works great. One of these days
I'll get around to packaging it up, and releasing it, if people are
interested.

bash$ wc -l ftpfs.py testfs.py
    171 ftpfs.py
     46 testfs.py
(the library is a little bigger! :-)

BTW: it doesn't use FUSE (or NFS - yucky yuk!, or even samba - yes
that's been done too!) - just standard Linux kernel modules...

However, ideally, I would prefer to be using Ron's v9fs with namespaces!

Martin

PS, my list of user-mode filesystem projects is:
	avfs/fuse: http://sourceforge.net/projects/avf/
	userfs: http://www.penguin.cz/~jim/userfs/
	uvfs: http://www.sciencething.org/geekthings/index.html
	Podfuk: http://atrey.karlin.mff.cuni.cz/~pavel/podfuk/podfuk.html
	lufs: http://lufs.sourceforge.net/lufs/
	virtualfs: http://www.solucorp.qc.ca/virtualfs/

	(although the last subverts libc, rather than using a kernel module)

--
Martin C. Atkins			martin@parvat.com
Parvat Infotech Private Limited		http://www.parvat.com{/,/martin}

Re: [9fans] user-level file systems for Linux

[ron minnich]

On Thu, 19 Feb 2004, Martin C.Atkins wrote:

> On Wed, 18 Feb 2004 20:57:42 -0500 "Russ Cox" <rsc@plan9.bell-labs.com> wrote:
> >...
> > Actually there have been a handful of projects doing
> > similar things over the past few years, and they've all
> > died out.  I'm optimistic about FUSE because it's almost
> > an exact translation of the VFS layer, meaning that it's
> > simple and as expressive as possible.
> >
> > Russ
>
> I wrote a user-mode filesystem for Linux just yesterday. In Python
> (using a library I hacked up early last year). Works great. One of these days
> I'll get around to packaging it up, and releasing it, if people are
> interested.

w.r.t. these things and 9p2000: it seems to me there are ops in the Linux
VFS that can not be expressed in 9p2000. Three that come to mind are
symlinks, hard links, and extended attributes (used for ACLs). The Linux
inode fields have some things that 9p2000 doesn't have, and vice-versa.

I had to add hard link/symlink ops to my 9p2000 to make things work as
expected on linux (else users will get mad). So I got several more ops for
things I really did not want to add, but had to.

I'm not sure I see how 9p2000 could support everything that FUSE would
want to do. Take a look inside any existingt NFS client/server to see how
really crazy it gets once named pipes, device nodes, etc. start to creep
in.

The goal of using 9p2000 with these things may require some difficult
decisions ... possibly the easiest thing to do is learn NFS V4 :-(

ron

Re: [9fans] user-level file systems for Linux

[Charles Forsyth]

>>I'm not sure I see how 9p2000 could support everything that FUSE would
>>want to do. Take a look inside any existingt NFS client/server to see how
>>really crazy it gets once named pipes, device nodes, etc. start to creep
>>in.

have they added a (guaranteed, non-optional) Close operation
to NFS?  otherwise, it's hard to see how NFS could be used
to distribute things in general.

Re: [9fans] user-level file systems for Linux

[ron minnich]

On Thu, 19 Feb 2004, Charles Forsyth wrote:

> have they added a (guaranteed, non-optional) Close operation
> to NFS?  otherwise, it's hard to see how NFS could be used
> to distribute things in general.

not sure.

ron

Re: [9fans] user-level file systems for Linux

[boyd, rounin]

ron minnich wrote:
>>have they added a (guaranteed, non-optional) Close operation
>>to NFS?  otherwise, it's hard to see how NFS could be used
>>to distribute things in general.
>
> not sure.

they had better had.  lack of close is a nightmare.

Re: [9fans] user-level file systems for Linux

[Stephen Wynne]

Yes, NFSv4 is stateful. See 1.4.4.  OPEN and CLOSE
at http://www.ietf.org/rfc/rfc3530.txt
Note that the equivalent of CIFS OPLocks is handled
by delegations.

It's a bit like AFS's callback mechanism. The Linux
kernel changes are still under development, see
http://www.citi.umich.edu/projects/nfsv4/.

More on V4 at http://www.nfsv4.org/

> they had better had.  lack of close is a nightmare.