Re: [9fans] Private Namespaces for Linux

[Ronald G Minnich <rminnich@lanl.gov>]

On Tuesday 20 November 2001 14:08, Alexander Viro wrote:

> > I put private namespaces in too. I.e. I put in system call to support
> > Plan-9 style session, attach, mount, and then have the VFS read/write
> > etc. go to a VFS with plan9-style operations, communicating 9p to a
> > server (I only have two: a memfs and a ufs). Union mounts work in the
> > plan 9 style. I also have a library which does the "hijack function calls
> > to libc" so that you have it from user mode on other OSes.
>
> Ugh.

Ugh? what's ugh? You can create a local name in your private name space that
connects to a remote place, with the remote server being accessed via 9p, and
the remote server being pretty much any 9p server. Which may or may not be
dishing out stuff from a file system or some other place (the
server-of-services list I mentioned in the paper, or other things). I don't
get your point.

> To start with, namespaces had been available on 2.4 since at least
> March.  See patches on ftp.math.psu.edu/pub/viro/ - namespaces-* is that
> + related VFS cleanups.  Currently most of that is merged into the main
> tree. Remaining stuff will go in when I'll somewhat clean the backlog on
> other things.  And it's a real thing - it isn't limited to #M combinations,
> any fs can be used.

Well, which kernel version. I've been playing Kernel-of-the-week here, and
there's still stuff that won't compile if you turn it on. I did try (but
failed) to use your patches. I'll try again, it looked good if I could get it
to work.

Also bear in mind I wrote this code in 1995-6 or so (so long ago I forget the
exact dates), when I still had to argue with people about why private name
spaces are better than 'chroot'. My first port was to 2.0, not 2.4. This
stuff had lain unused for 2.5 years until the Dobbs article appeared. I did
rebuild it and the user-mode stuff all works, which makes my day. Also I had
*zero* access to Plan 9 when I did it -- which is why I did it!

I'll look at your stuff again, if it's better than mine (probably) I'll use
it. I'm more concerned with having something to use, so if your stuff will do
what I want, I'm fine with that.


> BTW, could you use the standard meanings of terms?  Filesystem is
> filesystem, whether it's disk-based or completely syntetic.

I think I used the term 'server', which I prefer. I thought that was the Plan
9 usage anyway. I have found the term "filesystem" is a bad term, since so
many people interpret it so many different ways -- but I slip up and use it
anyway...

> VFS is more or
> less an equivalent of chan.c and sysfile.c - what you had described looks
> rather like devmnt.c with some strange additions.

Well, things get a bit mixed up when we discuss this stuff. There is not an
exact correspondance between bits of Linux and bits of Plan 9. But no, it is
not devmnt.c. I'm not quite sure why you think this is like the devmnt stuff.

ron