From mboxrd@z Thu Jan  1 00:00:00 1970
Message-ID: <e8b0dc29cb22e0958fac9086eabc9b49@vitanuova.com>
To: 9fans@cse.psu.edu
Subject: Re: [9fans] book chapters
From: rog@vitanuova.com
In-Reply-To: <3F006BC5.4070309@nas.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="US-ASCII"
Content-Transfer-Encoding: 7bit
Date: Mon, 30 Jun 2003 19:40:56 +0100
Topicbox-Message-UUID: e1f0ea8c-eacb-11e9-9e20-41e7f4b1d025

> Where I have problems with matt's solution breaking down is when you
> have worker_5 creating random sets of files in a traditional hierarchy.
>   Any subfolders he/she/it creates lose the rw permissions for the
> bosses group, and even with the sticky bit set for the group you lose
> those permissions the next level deep.

actually, you're thinking in unix terms there.  plan 9 has different
semantics, such that the attributes of a directory *are* inherited.

% cd /tmp
% mkdir x
% ls -ld x
d-rwxr-xr-x M 9 rog rog 0 Jun 30 18:57 x
% chgrp sys x
% chmod 770 x
% ls -ld x
d-rwxrwx--- M 9 rog sys 0 Jun 30 18:57 x
% mkdir x/foo
% ls -ld x/foo
d-rwxrwx--- M 9 rog sys 0 Jun 30 18:57 x/foo
% > x/foo/bar
% ls -l x/foo
--rw-rw---- M 9 rog sys 0 Jun 30 18:57 x/foo/bar
%

> why not just a write a file server that mounts itself on / and implements
> whatever policy you'd like to dream up?  it would just implement the
> policy, whatever that might be, and use the underlying filesystem to
> store the bits.
>
> isn't it just protocol conversion?

the problem is that if parts of that filesystem are bound elsewhere in
the namespace, it's difficult to find out how to access the attributes
of a particular file without knowing where it comes from.

i've sometimes thought about a system where one might be able to
access filesystem-specific meta-information on a file through that
file itself.

> I was wondering if you could even implement the data and resources forks
> with 9p. The file would be somefile, with the actual data as
> somefile/data
> and resource fork under somefile/resource

the problem with this is that *all* files have to be stored this way,
or...  you have to try and open "file", then "file/data", and reading
the parent directory (the container of "file") doesn't tell you
anything useful about the attributes of the files it contains (you
have to ls */data).

NeXTstep (and presumably now macos X) does something like this for
applications, but it never seemed like a particularly elegant
solution, as it's not transparent to code that really doesn't care
about this kind of thing.

to wander away from reality a little:

one could add a 9p message, say "walkmeta", similar to
walk, but which provides a fid that corresponds to the
meta-information for a particular file.

there is precedent for using a fid for out of band information: the
fid created by attach and used by auth is just there to negotiate
protocol meta-information.

of course, what the conventions might be on such a fid is open to
argument: does it just look like a file which can be read and written,
or could it be a directory hierarchy (allowing meta-meta information,
perhaps?  :-]).

to follow this flight of fancy a little further, suppose that walkmeta
does provide a fid which can be further walked, and we add a new path
separator, say "#", such that evaluating
	x#y
means "walk to x in the normal namespace; then walkmeta to y".

then one could imagine a scenario similar to the following:

% ls -l x
--rw-rw---- M 9 rog sys 0 Jun 30 18:57 x
% cat 'x#acl'
rog
foo
bar
% echo add baz > 'x#acl'
echo: write error: unkown user
% cd 'x#.'
% ls
acl
% pwd
x#.
% ls -l
acl
fs
resource
% ls -l 'acl#.'
ls: acl#.: no metadata
%

just musing.