Re: [RFC PATCH v2 00/13] Namespace for namespace roots
From: Andy Lutomirski
Date: Wed Nov 18 2026 - 12:46:30 EST
On Wed, Nov 18, 2026 at 9:08 AM Jarkko Sakkinen <jarkko@xxxxxxxxxx> wrote:
>
> On Wed, Nov 18, 2026 at 07:04:21PM +0200, Jarkko Sakkinen wrote:
> > On Wed, Nov 18, 2026 at 08:59:39AM -0800, Andy Lutomirski wrote:
> > > On Tue, Nov 17, 2026 at 3:14 PM Jarkko Sakkinen <jarkko@xxxxxxxxxx> wrote:
> > > >
> > > > OVERVIEW
> > > > ========
> > > >
> > > > A root namespace (rootns) is a kernel object backed by a file descriptor
> > > > that binds together:
> > > >
> > > > 1. A full or partial set of detached namespaces.
> > > > 2. A dedicated PID namespace for spawning processes into, signaling
> > > > members, and polling or waiting on PID 1 exit.
> > > > 3. An initially empty mount namespace (CLONE_EMPTY_MNTNS) that must be
> > > > populated before going live, without inheriting host mount-tree
> > > > ancestors.
> > > >
> > > > Work on this feature started in 2025 from David Howells' container object
> > > > patch set [1], which I turned into a namespace for namespace roots and
> > > > rebuilt around that premise.
> > > >
> > > > While the patch set is manually written, the original motivation came
> > > > from isolating LLM agents. Running untrusted agent workloads made it
> > > > clear that OCI runtimes need to minimize their live configuration
> > > > surface.
> > > >
> > > > Assembling mounts, credentials, and namespaces inside helper processes
> > > > with a half-configured container is what I see as the main challenge,
> > > > and the existential debate around namespaces and container objects is an
> > > > unfocused discussion topic.
> > >
> > >
> > > I've skimmed some of the series, and I see a bunch of good ideas that
> > > involve using an fd to manage a thing. But I'm wondering how much of
> > > the value is added by that and how much is added by having a new
> > > *type* of thing: the "rootns".
> > >
> > > How much of the value comes from the rootns vs the new APIs to
> > > configure namespaces without entering them first?
> > >
> > > For that matter, the cells series also seems to contain a (possibly
> > > very similar) variant of the same thing.
> > >
> > > Would it make sense to come up with a patchset that does *just* the
> > > configure-a-namespace-by-fd part? It seems to me that this should be
> > > an easier thing to agree on, and it's a big part of the complexity of
> > > all of this, and it solves a real problem even without
> > > cells/rootnses/jails/whatever.
> >
> > Is it more accurately scope the part that is "just the fd-part" and
> > what you consider as "extra cruft"?
I'm not actually trying to say that anything is "extra cruft". What I
mean is: there seems to be somewhat of a core that is pretty similar
between the two patchsets, and that core seems independently useful.
> >
> > I.e., where do you draw the border?
>
> I'm asking this because I had similar thoughts when I started this back
> in 2025. David had some good ideas but it did more than having a fd
> that does not go live before everything is set up properly.
I don't know that I have a very clear border, but here's an example:
rootns has MOVE_MOUNT_T_ROOTNS_ROOT.
cells has:
+What else goes into the tree is mounted on it before it becomes a
mount namespace. A
+filesystem for the cell to write to is mapped in the same way:
+
+.. code-block:: c
+
+ fs = fsopen("tmpfs", 0);
+ fsconfig(fs, FSCONFIG_SET_STRING, "mode", "1777", 0);
+ fsconfig(fs, FSCONFIG_CMD_CREATE, NULL, NULL, 0);
+ mnt = fsmount(fs, 0, MOUNT_ATTR_NOSUID | MOUNT_ATTR_NODEV);
+ mount_setattr(mnt, "", AT_EMPTY_PATH, &attr, sizeof(attr));
+ move_mount(mnt, "", tree, "tmp", MOVE_MOUNT_F_EMPTY_PATH);
(Source: https://lore.kernel.org/all/20261105-b4-cells-v3-2-bc9095937a84@xxxxxxxxxxxxxx/)
I feel like you are both *so close* to something that would stand
alone and make perfect sense. I also feel like, if that's a goal,
then both series are slightly off from the bulls-eye. (Keep in mind
that I've spend a total of about ten minutes looking at the relevant
parts of the series.)
If I've understood the two series correctly, and if I'm thinking of
these abilities as something that can stand alone, then I see the
following things that I'd improve.
rootns: Why can't I use a *mountns* fd to so the configuration? Maybe
I'm misunderstanding something.
cells: The dance where I create an anonymous mountns, mount stuff into
it, and then attach it to the cell is cute. But I think the userns
interaction is wrong: you are making me assemble all the mounts in
*my* userns and then clone the entire tree into a fresh, non-anonymous
mountns bound to the current userns. And that clone operation is not
free and serves no actual practical purpose.
Why can't there be some variant of these where I can either create an
anonymous mount tree that is genuinely unbound from a userns or that
lives in *the target userns* or that *is already the final mountns*,
using just mount-related APIs? In pseudocode:
int mountns_control_fd = create_mountns(either userns-free or with a
userns_control_fd);
fsopen, move_mount, , open_tree, etc;
Finally:
cells: I find it *incredibly annoying* that one generally cannot put
new mounts into an existing container using most existing container
managers. I hit this on a surprisingly regular basis. You seem to be
baking this into the whole cells design. Why?
both of you: Can we please make it possible to do the parts of this
that oughtn't require privilege without privilege? If the cap_userns
thing ever lands, cap_userns & CAP_SYS_ADMIN ought to be more than
sufficient to permit creating fresh mounts even without creating a
userns as long as we don't do anything that affects privileges within
our own userns. Or let us create a userns, *not enter it*, and do all
the mounting without actually having CAP_SYS_ADMIN in the permitted
set.
--
Andy Lutomirski
AMA Capital Management, LLC