On Tue, 2009-01-06 at 17:42 -0500, Trond Myklebust wrote:
> On Tue, 2009-01-06 at 15:58 -0600, Serge E. Hallyn wrote:
>
> > So should we use patch 2/4, plus (as someone - was it you? - suggested)
> > using a DEFAULT instead of init_utsname()->nodename when
> > current->utsname() == NULL?
>
> No. I'm don't think that 2/4 is correct either. Basically, 2/4 is saying
> that the container that first mounts the filesystem 'owns' it. However
> at the same time we know that the lifetime of the filesystem is in no
> way bounded by the lifetime of the container, and that's what gets you
> into trouble with 'umount' in the first place.
>
> IMO, the current code is the most correct approach, in that it assumes
> that the filesystems are owned by the 'init' namespace.
IMHO This seems more incorrect than trying to use a more proximal
namespace.
Cheers,
-Matt Helsley
On Tue, 2009-01-06 at 16:08 -0800, Matt Helsley wrote:
> IMHO This seems more incorrect than trying to use a more proximal
> namespace.
You have yet to explain why.
--
Trond Myklebust
Linux NFS client maintainer
NetApp
[email protected]
http://www.netapp.com
On Tue, Jan 06, 2009 at 04:08:50PM -0800, Matt Helsley wrote:
> On Tue, 2009-01-06 at 17:42 -0500, Trond Myklebust wrote:
> > On Tue, 2009-01-06 at 15:58 -0600, Serge E. Hallyn wrote:
> >
> > > So should we use patch 2/4, plus (as someone - was it you? - suggested)
> > > using a DEFAULT instead of init_utsname()->nodename when
> > > current->utsname() == NULL?
> >
> > No. I'm don't think that 2/4 is correct either. Basically, 2/4 is saying
> > that the container that first mounts the filesystem 'owns' it. However
> > at the same time we know that the lifetime of the filesystem is in no
> > way bounded by the lifetime of the container, and that's what gets you
> > into trouble with 'umount' in the first place.
> >
> > IMO, the current code is the most correct approach, in that it assumes
> > that the filesystems are owned by the 'init' namespace.
>
> IMHO This seems more incorrect than trying to use a more proximal
> namespace.
If it would be possible, for example, for the 'init' namespace to have
no network interfaces at all, then it would be nicer to use a name
that's at least been used with nfs at *some* point--just on the general
principle of not leaking information to a domain that the user wouldn't
expect it to.
(Assuming it's unlikely anyone would consider init's utsname to be
sensitive information, that's a minor point.)
--b.
On Tue, 2009-01-06 at 19:20 -0500, J. Bruce Fields wrote:
> If it would be possible, for example, for the 'init' namespace to have
> no network interfaces at all, then it would be nicer to use a name
> that's at least been used with nfs at *some* point--just on the general
> principle of not leaking information to a domain that the user wouldn't
> expect it to.
Then RPC would fail. Thanks to the limitations imposed by selinux &
friends, all RPC sockets have to be owned by the init process.
--
Trond Myklebust
Linux NFS client maintainer
NetApp
[email protected]
http://www.netapp.com