overthinkos/container-nesting

container-nesting -- Rootless nested podman, buildah, skopeo

Overview

Adds everything needed to run rootless podman/buildah/skopeo inside a rootless outer container — at the default uid 1000, with zero added capabilities, no --privileged, no seccomp=unconfined, no label=disable. The recipe is a direct port of quay.io/podman/stable's canonical configuration, ported into the charly candy system so any box can compose it.

Candy Properties

| Property | Value | |----------|-------| | cap_add | (none) | | security_opt | unmask=/proc/* | | devices | /dev/fuse, /dev/net/tun | | Volumes | storage at /var/lib/containers/storage (only used by root images) | | Env | CHARLY_BUILD_ENGINE=podman, CHARLY_RUN_ENGINE=podman, _CONTAINERS_USERNS_CONFIGURED="", BUILDAH_ISOLATION=chroot |

Packages

RPM: buildah, fuse-overlayfs, shadow-utils, skopeo, tailscale, libsecret (Tailscale from the tailscale-stable repo).

Pacman: buildah, crun, fuse-overlayfs, libsecret, podman, shadow, skopeo, tailscale.

Arch must declare podman and crun explicitly: the candy's whole purpose is rootless nested podman, and the containers.conf shipped by this candy explicitly sets runtime = "crun". RPM users get podman transitively via the Fedora base image; Arch has no such transitive pull, so both podman and crun are declared explicitly in the pac: list (declaring docker instead, or omitting crun, leaves the Arch-based box with no podman binary in $PATH).

The kernel-level RCA (why none of the obvious fixes work)

This is the load-bearing section — if you're here because "nested podman fails with crun: mount proc to proc: Operation not permitted", read this before trying anything else.

What actually fails

When the inner podman starts a new container (e.g., alpine) inside a rootless outer container, crun tries to mount a fresh procfs for the new container's mount namespace. The call is roughly:

mount("proc", "/proc", "proc", MS_NOSUID|MS_NODEV|MS_NOEXEC, NULL);

The Linux kernel refuses with EPERM. Not because of capabilities, not because of seccomp, not because of SELinux.

Why the kernel refuses

fs/namespace.c:mount_too_revealing() is a security check introduced to prevent information leakage across user-namespace boundaries. It rejects a procfs mount when:

• The calling process is in a descendant user namespace of the existing procfs's owning user namespace, AND
• The existing procfs has submounts that the new mount would expose (the classic example: /proc/kcore has been bind-mounted over with /dev/null to hide kernel memory, and a fresh mount would "un-hide" it).

Podman's default rootless outer container generates an OCI spec with linux.maskedPaths covering:

/sys/kernel, /proc/acpi, /proc/kcore, /proc/keys,
/proc/latency_stats, /proc/sched_debug, /proc/scsi,
/proc/timer_list, /proc/timer_stats,
/sys/devices/virtual/powercap, /sys/firmware,
/sys/fs/selinux, /proc/interrupts

Each of these paths is either bind-mounted over with /dev/null or mounted as a read-only tmpfs. When the inner container tries to mount its own fresh /proc, the kernel sees that would reveal those paths → mount_too_revealing → EPERM.

Why capability-based fixes don't work

Empirically tested:

| Attempt | Result | |---|---| | --cap-add=SYS_ADMIN | FAIL — caps aren't the issue | | --cap-add=ALL | FAIL — same | | --cap-add=ALL --security-opt seccomp=unconfined --security-opt label=disable | FAIL — caps + seccomp + SELinux aren't the issue | | --privileged | PASS — but only because --privileged coincidentally also removes the masked_paths | | --security-opt unmask=/proc/* | PASS — surgical fix, no caps needed |

unmask=/proc/* tells podman NOT to emit those maskedPaths entries for /proc on the outer container. With nothing to mismatch, mount_too_revealing has nothing to reject. The inner /proc mount proceeds cleanly.

This candy's security: block is security_opt: [unmask=/proc/*] + devices: [/dev/fuse, /dev/net/tun]. No capability added. No seccomp touched. No SELinux touched. The surgical minimum.

Security trade-off

unmask=/proc/* exposes /proc/kcore (kernel memory) and /proc/keys (kernel keyring) on the outer container's filesystem.

Reading those files still requires CAP_SYS_ADMIN in the init user namespace — which a rootless container never has. The actual information leak is minimal. Compared to --privileged (which ALSO removes the masks, plus grants every capability, plus disables seccomp, plus passes through every host device, plus disables path masking entirely), this is the least-privilege fix available.

Subuid / subgid layout (must fit inside the outer namespace)

charly shell launches the outer container with --userns=keep-id:uid=1000,gid=1000 (default — see charly/shell.go:254). That creates a uid_map inside the outer of:

0    1000    1           # inner uid 0 → host uid 1000
1    100000  65535       # inner uid 1-65535 → host uid 100000-165534

So inside the outer, only inner uids 0-65535 exist. Subid delegation ranges that fall outside this window fail at newuidmap write to uid_map: EPERM.

The candy emits two non-overlapping ranges for the primary uid-1000 user, skipping uid 1000 itself (because keep-id already owns it):

user:1:999
user:1001:64535

…plus a full-range entry for root (used by charly-fedora/charly-arch/ githubrunner, which run as uid 0):

root:1:65535

This pattern matches quay.io/podman/stable's /etc/subuid layout exactly. A range like 524288:65536 would fall outside the outer namespace's mapped window and cause an obscure newuidmap write failure — the delegation ranges MUST fit inside the keep-id window.

The newuidmap/newgidmap binaries get cap_setuid=ep / cap_setgid=ep file capabilities (via a dedicated task) so any uid invoking them can delegate subids.

The setcap(8) binary MUST be installed for that task to do anything. Fedora and Arch ship it transitively (libcap is in the base image), but Debian/Ubuntu do NOT — so the candy's deb sections declare libcap2-bin explicitly. Without it the setcap cap_setuid=ep /usr/bin/newuidmap step exits 127 (setcap: command not found) and, because the step is a bare RUN, the build hard-fails on the deb path (and any image that did slip through would ship a capability-less newuidmap, so nested podman dies at runtime with newuidmap: open of uid_map failed: Permission denied). RDD-confirmed 2026-06-15: this is the one non-format-agnostic deb dependency — every other piece of the recipe is shared.

Config files — written to both system-wide and user locations

Rootless podman prefers ~/.config/containers/* over /etc/containers/*. Writing only the system-wide location is a no-op for the desktop user. This candy writes every config to both locations.

containers.conf

Identical at /etc/containers/containers.conf and ~/.config/containers/containers.conf:

[containers]
cgroups = "disabled"
cgroupns = "host"
ipcns = "host"
netns = "host"
userns = "host"
utsns = "host"
log_driver = "k8s-file"

[engine]
cgroup_manager = "cgroupfs"
events_logger = "file"
runtime = "crun"

Why each setting:

• cgroups = "disabled" — rootless cgroupv2 delegation isn't guaranteed at nesting depth 2+; disabling avoids "cannot set up cgroup" errors.
• cgroup_manager = "cgroupfs" — systemd cgroup delegation likewise not guaranteed.
• netns = "host" — pasta (rootless networking) needs /proc/sys/net/ipv4/ping_group_range writable, which is read-only in a rootless outer. netns=host makes the inner reuse the outer's netns, bypassing the need.
• userns = "host" — required by the mount_too_revealing analysis above. Without it, the inner podman creates a descendant userns and hits the kernel check on every /proc mount.
• ipcns, utsns, cgroupns = "host" — match the canonical quay.io/podman/stable config; prevents a cascade of namespace permission errors observed when only some are set.

storage.conf

System-wide (for root images) at /etc/containers/storage.conf:

[storage]
driver = "overlay"
runroot = "/run/containers/storage"
graphroot = "/var/lib/containers/storage"

[storage.options.overlay]
mount_program = "/usr/bin/fuse-overlayfs"
mountopt = "nodev,fsync=0"

User-level (for uid-1000 images) at ~/.config/containers/storage.conf — different graphroot so it's user-writable:

[storage]
driver = "overlay"
runroot = "${HOME}/.local/share/containers/run"
graphroot = "${HOME}/.local/share/containers/storage"

[storage.options.overlay]
mount_program = "/usr/bin/fuse-overlayfs"
mountopt = "nodev,fsync=0"

mount_program = "/usr/bin/fuse-overlayfs" is the critical line — the kernel overlay driver can't mount from a rootless outer (no CAP_SYS_ADMIN in init userns); fuse-overlayfs can.

If the user-level file is missing, rootless podman uses the system default, which points at /var/lib/containers/storage — unwritable by uid 1000 → mkdir graphroot: permission denied on first podman run.

policy.json

Same content at both /etc/containers/policy.json and ~/.config/containers/policy.json:

{"default":[{"type":"insecureAcceptAnything"}]}

Without this, podman pull fails with no policy.json file found.

Env vars — the two hidden contracts

| Env var | Value | Role | |---|---|---| | _CONTAINERS_USERNS_CONFIGURED | "" (empty string, SET not UNSET) | Tells the inner podman "you're already inside a rootless user namespace". Without this, the inner re-execs itself via newuidmap to create a new descendant user namespace — defeating userns=host in containers.conf and re-triggering mount_too_revealing. | | BUILDAH_ISOLATION | chroot | Tells buildah RUN steps to use chroot isolation instead of the OCI runtime. Without this, nested podman build falls back to OCI isolation which creates a descendant user namespace and hits the same kernel check. |

Both are baked into the candy's env: section so they land in the OCI env of any box composing this candy.

Box-level compatibility (union semantics)

charly/security.go:66-97 unions box-level CapAdd, SecurityOpt, Devices onto the candy-level merged set (via appendUnique). Box values can only ADD, never strip.

Consequence: boxes that want the old full-hammer posture (charly-fedora, charly-arch, githubrunner) must assert it at the box level, not expect this candy to donate it. Their charly.yml entries carry:

security:
  cap_add: [ALL]
  security_opt:
    - label=disable
    - seccomp=unconfined

The resolved OCI label then unions to cap_add:[ALL] + security_opt:[unmask=/proc/*, label=disable, seccomp=unconfined], which matches their historical posture.

Rootless boxes like /charly-openclaw:openclaw-desktop don't add a box-level security: block, so the resolved posture stays at security_opt:[unmask=/proc/*] only — zero capability escalation.

Cross-distro coverage

rpm: (Fedora), pac: (Arch), plus per-distro deb sections (debian: / debian-13: / ubuntu: / ubuntu-24.04:) over the shared deb package set (podman, buildah, skopeo, fuse-overlayfs, crun, uidmap, passwd, libsecret-1-0, plus libcap2-bin — the setcap provider, see the subuid/subgid section above). The version-specific debian-13: / ubuntu-24.04: sections additionally declare the Tailscale apt repo (https://pkgs.tailscale.com/stable/debian or .../ubuntu) with signed-by key for the tailscale package. Drops on deb: none at the nesting recipe level — all the critical pieces (containers.conf, storage.conf, subuid layout, _CONTAINERS_USERNS_CONFIGURED="" env) are format-agnostic task content; the sole deb-only addition is libcap2-bin.

Usage — rootless image (uid 1000)

# charly.yml
openclaw-desktop:
  base: cachyos.cachyos
  candy:
    - selkies-desktop
    - openclaw-full
    - ollama
    - charly
    - container-nesting   # donates unmask + devices + config + env
    - ...
  # NO uid/gid/user/network override

Usage — root image (uid 0)

# charly.yml
charly-fedora:
  base: fedora
  uid: 0
  gid: 0
  user: root
  network: host
  security:
    cap_add: [ALL]
    security_opt:
      - label=disable
      - seccomp=unconfined
  candy:
    - charly
    - container-nesting
    - ...

Both paths work; they just resolve to different OCI security labels.

Verification

# Rootless posture on openclaw-desktop
charly box inspect openclaw-desktop | jq '.HostConfig? // .Config.Labels."ai.opencharly.security"'
# → cap_add:[], security_opt:[unmask=/proc/*], devices:[/dev/fuse,/dev/net/tun]

# Nested podman smoke (inside the running container)
charly shell openclaw-desktop -c 'podman run --rm quay.io/libpod/alpine:latest true'
# → exit 0, NO "mount proc to proc: Operation not permitted"

# Diagnostic: inspect the OCI spec generated for a nested container
charly shell openclaw-desktop -c '
  podman create --name t quay.io/libpod/alpine:latest /bin/true >/dev/null
  sf=$(find ~/.local/share/containers -name config.json -path "*/userdata/*" | head -1)
  jq ".linux.maskedPaths" "$sf"'
# → empty list or no /proc entries = unmask worked

If mount proc to proc: EPERM still happens after a rebuild, check in this order:

1. env | grep _CONTAINERS_USERNS_CONFIGURED — must print one line with empty value (SET, not UNSET).
2. grep '^userns' /etc/containers/containers.conf ~/.config/containers/containers.conf — both must say host.
3. cat /etc/subuid — must show the 1:999 + 1001:64535 pattern for the primary user.
4. podman inspect <outer-container> --format '{{.HostConfig.SecurityOpt}}' — must include unmask=/proc/*.

Used In Boxes

• /charly-openclaw:openclaw-desktop — rootless path; box-level adds nothing
• /charly-distros:charly-fedora — root path; box-level adds cap_add:[ALL] + security_opt:[label=disable, seccomp=unconfined]
• /charly-coder:charly-arch — same root path as charly-fedora
• /charly-distros:githubrunner — same root path; doesn't compose the full charly toolchain but keeps nested podman for CI workloads

Related Candies

• /charly-tools:charly — pairs with container-nesting in charly-toolchain images (the full toolchain: charly binary + VM + encrypted storage tools)
• /charly-infrastructure:virtualization — supervisord-managed rootless libvirt (virtqemud, virtnetworkd). Pairs with container-nesting for images that need both nested containers AND nested VMs
• /charly-coder:sshd — sibling enabling remote access to nested-container hosts

Related Commands

• /charly-build:build — build images that ship nested podman
• /charly-core:shell — run nested podman/buildah commands inside the outer
• /charly-build:generate — Containerfile generation (the service: supervisord fragments for container-nesting consumers go through the fragment_assembly init model)
• /charly-core:charly-config — box-level security: union with candy-level when deploying

When to Use This Skill

MUST be invoked when:

• Authoring or debugging any candy/box that needs nested podman, buildah, or skopeo.
• Chasing mount_too_revealing / mount proc to proc: Operation not permitted errors — this is the authoritative RCA.
• Choosing between --privileged, cap_add: ALL, and unmask=/proc/* — this skill documents why the surgical unmask fix is the minimum-privilege path and the others are hammers.
• Evaluating the security posture of /charly-openclaw:openclaw-desktop, /charly-distros:charly-fedora, /charly-coder:charly-arch, or /charly-distros:githubrunner.

Related

• /charly-image:layer — candy authoring reference (charly.yml schema, task verbs, service declarations)
• /charly-check:check — declarative testing (check: block, charly check box, charly check live)