On Wed, Nov 18, 2026 at 09:35:26PM +0200, Jarkko Sakkinen wrote:
OVERVIEW
========
A root namespace (rootns) is a kernel object backed by a file descriptor
that binds together:
1. A full set of detached namespaces owned by a fresh user namespace.
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) populated
directly in its target user namespace before going live, without
inheriting host mounts or cloning staged trees, while still allowing
new mounts to be attached at runtime.
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.
[1] https://lore.kernel.org/all/155059610368.17079.2220554006494174417.stgit@warthog.procyon.org.uk/ (local)
IMPLEMENTATION
==============
rootns_create() allocates a rootns instance backed by a pollable
anonymous inode fd, creating a fresh owning user namespace and a full set
of detached namespaces: empty mount (CLONE_EMPTY_MNTNS), PID, UTS, IPC,
network, cgroup, and time. ioctl(rootnsfd, ROOTNS_GET_NSFD, nstype)
returns nsfs file descriptors for individual member namespaces when
standalone namespace operations are needed.
Superblocks are configured directly in the target user namespace via
fsopen(), FSCONFIG_SET_ROOTNS, and fsmount(), allowing unprivileged
callers to assemble FS_USERNS_MOUNT filesystems without holding
CAP_SYS_ADMIN in their own user namespace or entering the target user
namespace first. move_mount(..., MOVE_MOUNT_T_MNTNS) operates on any
non-anonymous mount namespace fd: on an empty mount namespace it installs
"/" onto the nullfs root once, and it attaches relative submounts both
before entry and into live mount namespaces.
Once "/" is installed, processes are spawned into the rootns via clone3()
with CLONE_INTO_ROOTNS and struct clone_args.rootns, without mutating the
caller in place. Pairing CLONE_INTO_ROOTNS with CLONE_VM | CLONE_VFORK
and opening the target binary via openat2(..., RESOLVE_IN_ROOT) on the
retained "/" mount fd allows spawning with execveat() without copying
page tables.
The first entrant becomes PID 1; once it exits and
zap_pid_ns_processes() reaps all members, poll() reports EPOLLHUP and
rootns_wait() returns init's exit status and rusage. rootns_kill() walks
pid_ns->idr under tasklist_lock to signal member thread groups.
CHANGES
=======
v3:
- The patch set size increased predominantly because the coverage of
kselftest had to widened to cover the introduced changes.
- Replace MOVE_MOUNT_T_ROOTNS_ROOT with MOVE_MOUNT_T_MNTNS targeting
mnt_ns nsfs fds, and add ioctl(ROOTNS_GET_NSFD) to open member
namespace fds (Andy Lutomirski).
- Allow unprivileged fsopen(), FSCONFIG_SET_ROOTNS, and fsmount() to
configure and mount FS_USERNS_MOUNT filesystems directly in the target
user namespace without entering it first (Andy Lutomirski).
- Fix locking, refcounting, credential/dumpability/timens setup,
pivot_root() lookup, and cross-userns mount checks across copy_rootns(),
CLONE_INTO_ROOTNS, rootns_kill(), FSCONFIG_SET_ROOTNS, and move_mount()
(Sashiko).
- Mark PowerPC rootns syscalls nospu (Sashiko).
- Update documentation and kselftests.
I.e. before PID 1 and /proc/<pid>/ns/* existss, nsfd for namespace other
than root nsfd can be acquired as follows:
int mnts_fd = ioctl(root_fd, ROOTNS_GET_NSFD, CLONE_NEWNS);
Br, Jarkko