Re: [RFC PATCH v2 00/13] Namespace for namespace roots
From: Andy Lutomirski
Date: Fri Nov 20 2026 - 14:53:50 EST
On Fri, Nov 20, 2026 at 11:48 AM Jarkko Sakkinen <jarkko@xxxxxxxxxx> wrote:
>
> On Fri, Nov 20, 2026 at 09:35:06PM +0200, Jarkko Sakkinen wrote:
> > On Fri, Nov 20, 2026 at 03:57:32PM +0200, Jarkko Sakkinen wrote:
> > > On Fri, Nov 20, 2026 at 02:24:57PM +0100, Alexander Mikhalitsyn wrote:
> > > > Yeah, detached mounts API solves a security issue here, but it does not
> > > > fully enable the use-case Andy describes (if I understood it
> > > > correctly). My interpretation here is that to fully cover it we need
> > > > some well-isolated communication channel between container and
> > > > container manager (vsock for containers? :-D). And some well-defined
> > > > agreement of how PID 1 inside the container can get an information
> > > > about a hotplugg-ed stuff. [ But it could be that I completely
> > > > misunderstood and Andy meant something much more simple here. ]
> >
> > The sequence for hotplug (Andy's use case) breaks down like this and is
> > doable:
> >
> > 1. Host creates rootns(ROOTNS_SEAL_MNTNS).
> > 2. Host sets uid/gid maps with NS_USERNS_SET_IDMAP.
> > 3. Host populates "/" with MOVE_MOUNT_T_MNTNS, optionally combined with
> > MOVE_MOUNT_CLONE for bind mounts.
> > 4. Host spawns PID 1 with clone3(CLONE_INTO_ROOTNS) (optionally passing
> > an AF_UNIX socketpair fd for communication).
> > 5. PID 1 watches /proc/self/ns/mnt via fanotify(FAN_REPORT_MNT |
> > FAN_MARK_MNTNS, FAN_MNT_ATTACH) or polls /proc/self/mountinfo.
> > 6. On hotplug, host mounts the filesystem via fsopen()/fsmount() (plus
> > optional MOUNT_ATTR_IDMAP).
> > 7. Host creates the mountpoint with openat2(RESOLVE_IN_ROOT) + mkdirat()
> > and attaches the mount with MOVE_MOUNT_T_MNTNS.
> > 8. PID 1 receives FAN_MNT_ATTACH and queries the new mount via
> > statmount().
> >
> > ROOTNS_SEAL_MNTS, NS_USERNS_SET_IDMAP and MOVE_MOUNT_CLONE new features
> > that are yet to be submitted but they were not tedious to ramp up.
> >
> > I'll include them to v5 of the patch set. Thanks for underlining this
> > part because that led to having those.
> >
> > Architecture scales to this even if my choices to address these details
> > would need some adjustment.
> >
> > The clear two-phase (build/live) progress with rootns sort of just gives
> > room to address stuff like this, and I have hard time to believe that it
> > would get much better than this.
>
> A completely new feature is NS_USERNS_SET_IDMAP:
>
> nsfs: Support NS_USERNS_SET_IDMAP ioctl
>
> Configuring UID, GID, and project ID mappings or disabling setgroups(2)
> for a user namespace currently requires writing to /proc/<pid>/uid_map,
> gid_map, projid_map, and setgroups, forcing container managers to spawn a
> live process before populating a filesystem owned by that user namespace.
>
> Add the extensible NS_USERNS_SET_IDMAP ioctl on CLONE_NEWUSER nsfs file
> descriptors to atomically configure setgroups policy and UID, GID, and
> project ID mappings before any process enters the namespace. Enforce the
> same single-write and capability invariants against both current_cred()
> and filp->f_cred under userns_state_mutex.
>
> Signed-off-by: Jarkko Sakkinen <jarkko@xxxxxxxxxx>
>
I'm highly in favor of this. But maybe not as an ioctl.
If you have an nsfd for a userns and that nsfd is the "control" fd
(however anyone decides this is formulated), why does f_cred matter?
It should be possible to create a userns without privilege and
SCM_RIGHTS the fd to a privileged uidmapd. (I fixed the rather
screwed up workflow for this in newuidmap/newgidmap a few weeks ago --
this is remarkably easy to screw up. It'll make it to a distro near
you some day. It's supposed to *work* even without CAP_SYS_ADMIN in
the picture, and it did not.)
--Andy
--
Andy Lutomirski
AMA Capital Management, LLC