Re: [PATCH -mm 5/7] add user namespace

[ebiederm@xmission.com (Eric W. Biederman)]

Trond Myklebust <trond.myklebust@fys.uio.no> writes:

> On Fri, 2006-07-14 at 11:36 -0600, Eric W. Biederman wrote:
>> Dave Hansen <haveblue@us.ibm.com> writes:
>> 
>> > On Fri, 2006-07-14 at 12:08 -0500, Serge E. Hallyn wrote:
>> >> yes, of course, vfsmount, which I assume is what Eric meant?
>> >> 
>> >> Which means we'd have to do this at permission() using the nameidata, or
>> >> pass nd to generic_permission. 
>> >
>> > Yeah, I think so.  But, this is well into Al territory, and there might
>> > be a better way.
>> 
>> Well until we get that sorted out I will keep picking on i_sb.
>
> Don't bother: labelling superblocks with process-specific data is always
> going to be unacceptable.

It's not process specific data.  It is a pointer to global context in
which uid's on the filesystem uniquely specify a user.  This is
something that would get set when the filesystem is mounted.

> In order to avoid aliased superblocks, you would have to be able
> guarantee to be the sole owner of the data on the device that it refers
> to. You'd have to own the device in order to do that, in which case you
> are better off just labelling the device instead.

Now I do agree if I can set the information in vfsmount and not in
the superblock it is probably better.  But even with nfs mount superblock
collapsing (which I almost understand) I don't see it as a real
problem, as long as I could prevent the superblock from collapsing.

Eric