linux-container: Remount / recursively as MS_PRIVATE, immediately after entering the new mount namespace.

Apply the same workaround that runc, crun, and buildah apply
unconditionally: immediately after entering the new mount namespace,
remount / recursively as MS_PRIVATE. Concretely, the call we will
make in Guile mirrors the verbatim C call already present in crun at
https://github.com/containers/crun/blob/99a8512ef50c5af1b8cfff28eacddf870220ca9f/src/libcrun/linux.c#L4677:

    ret = mount (NULL, "/", NULL, MS_REC | MS_PRIVATE, NULL);

and the Go equivalents in:

- runc (defaults to MS_SLAVE for reasons documented in its pivotRoot):
https://github.com/opencontainers/runc/blob/a7e766484c38be1f0cf9b7d960c9324316cb7204/libcontainer/rootfs_linux.go#L1106

- buildah (expands to mount("", "/", "none", MS_REC|MS_PRIVATE, "")):
https://github.com/podman-container-tools/buildah/blob/769d311d739dd4bbde496a9480879ce6aadaa1e0/bind/mount.go#L41

Our implementation:

1. Detaches the container's mount namespace from the host's propagation
peer group, so subsequent mount/umount events inside the container
do not propagate to the host.
2. Makes the kernel's pivot_root(2) preconditions satisfied: both
new_root (the tmpfs we then create) and its parent (/) have propagation
type MS_PRIVATE.
3. Should have no effect on processes outside the new mount namespace. The
host's / retains whatever propagation type it had.

mount-file-systems is only invoked from run-container when both
'mnt namespace is requested and root != "/", so we are
guaranteed to be inside a fresh mount namespace at this point and that
the remount only affects this namespace. The remount must happen before
the tmpfs is created. Once the tmpfs exists as a child of the (shared)
host /, fixing only the host / does not retroactively change the
propagation type the tmpfs already inherited. Doing the remount first
means every subsequent mount this function performs inherits MS_PRIVATE
from its parent.

* guix/build/syscalls.scm (MS_PRIVATE): New variable.
* gnu/build/linux-container.scm (mount-file-systems): Remount / recursively as MS_PRIVATE, immediately after entering the new mount namespace.
* gnu/services/containers.scm (%miniflux-create-admin-credentials,
%rootless-podman-os-with-least-authority-wrapper,
run-rootless-podman-test-with-least-authority-wrapper): New bindings.
(%test-rootless-podman-with-least-authority-wrapper): New regression
test making sure the rootless-podman-service-type does not break again
least-authority-wrapper bases services.

Fixes: https://codeberg.org/guix/guix/issues/3233
Merges: https://codeberg.org/guix/guix/pulls/9140

1 files changed, 11 insertions, 0 deletions

diff --git a/gnu/build/linux-container.scm b/gnu/build/linux-container.scm
index bc1f8851dd2..32990f4dac5 100644
--- a/gnu/build/linux-container.scm
+++ b/gnu/build/linux-container.scm
@@ -1,6 +1,7 @@
 ;;; GNU Guix --- Functional package management for GNU
 ;;; Copyright © 2015 David Thompson <davet@gnu.org>
 ;;; Copyright © 2017-2019, 2022-2023, 2025 Ludovic Courtès <ludo@gnu.org>
+;;; Copyright © 2026 Giacomo Leidi <therewasa@fishinthecalculator.me>
 ;;;
 ;;; This file is part of GNU Guix.
 ;;;
@@ -104,6 +105,16 @@ for the process."
     (mkdir-p target)
     (mount source target type flags options #:update-mtab? update-mtab?))
 
+  ;; Detach our mount namespace from the host's propagation peer
+  ;; group.  Without this, when the host's / is MS_SHARED - as it
+  ;; is on a system running rootless-podman-service-type, and as it
+  ;; is by default on systemd hosts - 'pivot_root' below fails
+  ;; with EINVAL because the new root and its parent must not have
+  ;; propagation type MS_SHARED (pivot_root(2)).  This is the same
+  ;; remount that runc, crun and buildah perform immediately after
+  ;; unshare(CLONE_NEWNS).
+  (mount #f "/" #f (logior MS_REC MS_PRIVATE))
+
   ;; The container's file system is completely ephemeral, sans directories
   ;; bind-mounted from the host.
   (mount "none" root "tmpfs")