Updates

It is possible to rebuild a running system with vpsadminos-rebuild tool. The tool builds /etc/vpsadminos#$(hostname) by default, using the vpsadminosConfigurations.<hostname> flake output generated by vpsadminos-generate-config.

On an installed system, edit /etc/vpsadminos/configuration.nix and run:

vpsadminos-rebuild switch

To select a different configuration or flake path, pass --flake:

vpsadminos-rebuild switch --flake /etc/vpsadminos#my-node

You can also build the tool from a checkout:

make vpsadminos-rebuild

Running impermanent containers

Older osctld versions can save the base dataset and a disabled temporary-dataset cleanup flag before selecting an impermanent NixOS container's live temporary root. The container can continue running with correct in-memory state until the daemon is replaced. Current starts save the completed selection before the workload starts, independently of CPU topology or scheduler assignment.

During pool import, osctld repairs this saved base selection only when the complete saved run ID belongs to that pool and container, exactly one existing GC record owns the same run's temporary dataset, and no competing record names that dataset or an ancestor or descendant. It authenticates the retained live init process, its namespace PID 1, birth identity, cgroup, mount namespace and root. The ZFS source, GUID and host private directory must match that root. The checks run again before publication. The repair atomically updates only the saved dataset and cleanup flag, preserving the run ID, other saved fields and GC records. Already-correct impermanent selections and supported .boot-<suffix> selections with cleanup enabled undergo the same checks without rewriting YAML.

Enabling impermanence while a NixOS container is running takes effect on its next start. During import, osctld preserves an ordinary base-root run when the saved selection names the base with cleanup disabled, the held live process and root match that base, and GC ownership does not conflict. The GC file may be absent only when in-memory ownership is empty and both remain unchanged through the final checks. This case leaves the saved runtime configuration and GC records unchanged.

For temporary-root recovery, missing, empty, malformed, foreign, conflicting, ambiguous or changed evidence prevents the container from entering management. The daemon reports container omitted and unmanaged and inhibits temporary-dataset reclamation for that pool instance. This refusal preserves the live workload and storage; it does not reconstruct ownership from dataset names or host mount prefixes. If monitoring, console attachment or event reporting fails after publication, the same run remains managed and the daemon reports attachment incomplete; published container remains managed. Reclamation is inhibited in that case as well.

While reclamation is inhibited, periodic and manual pruning and queued explicit frees cannot move temporary datasets to the trash bin. Registered ownership records remain intact. Queue handling, new registrations, saving and collector shutdown still work. Datasets already in the trash bin are outside this safeguard, and a refused free without a registered ownership record has no durable automatic retry guarantee. Only a fresh import that loads all containers successfully resumes reclamation; there is no manual hold-release command. Preserve the reported runtime and GC metadata when investigating a refusal, and do not delete a live root to clear the error.

The writer and import repair use the existing YAML fields and run ID format; they add no API or schema version. Nodes can update independently without a coordinated fleet update. Updating a writer does not correct state already saved by an older daemon, so retaining the authenticated import reader matters during mixed-version operation. Older readers can parse repaired metadata, but rolling back osctld loses the authenticated recovery and reclamation safeguards. A userspace update preserves the running kernel and existing container mounts.

Selectable Linux 6.18 platform

The 6.18 release line selects Linux 6.18.49 with its matching OpenZFS and LXC sources. The configured stable kernel remains the default; moving to 6.18 is explicit:

system.vpsadminos.enableUnstable = true;

For a source-checkout build, os/configs/unstable.nix sets the same option. Changing the boot kernel requires a reboot; vpsadminos-rebuild switch alone does not replace the running kernel. Keep the previous bootable system generation until the new kernel and workloads have been verified. To return to the default kernel, disable this option, rebuild, and reboot into the resulting generation. The configured default and retained source entries are listed in os/packages/linux/available-kernels.nix. Keep complete kernel and OpenZFS tuples when maintaining rollback generations.

The selected source revisions are Linux a770f5086fc500ca258852aea1cb3e209a342d2b, OpenZFS 60536e063336cb5daaec84637297f738a399c44c, and LXC 267e7844c122b9be56ca8a1ccc3d2f4d8986c326. Kernel and OpenZFS selection is paired in os/packages/linux/available-kernels.nix; do not mix their modules with another kernel build.

The OpenZFS 2.3.x fork includes selected stability and security backports, including host-namespace authorization for pool administration and fault injection (CVE-2026-79619), vdev device-access checks, bounded nvlist decoding, and receive, rollback, lifetime, and MMP corrections. It is not a migration to OpenZFS 2.4.x or an automatic pool-feature upgrade. LXC remains the same 6.0.6-based downstream source; its package version now reflects that base. The rsyslog package also carries the narrow RainerScript replace() sizing repair for CVE-2026-78002 without changing its release line or defaults.

This line includes proxy execution, proactive-reclaim support, namespaced container tracing, and the corresponding container lifecycle and filesystem corrections. PSI and SCHED_CLASS_EXT remain disabled, as does host-global unprivileged BPF. The downstream authority and credential hardening families are excluded; ordinary Linux credentials, capabilities, LSMs and container permission checks remain. SELinux namespace/outer-context work and separate hardening development are not part of this release.

Existing containers and namespaced tracing

The namespaced-tracing additions are a feature of the selected platform, not a prerequisite for an ordinary osctld or system-generation activation. Keep the matching kernel/LXC support, private host tracing mounts and container bpffs lifecycle together; they do not enable host-global unprivileged BPF.

Containers that are already running keep their current namespaces and mounts when osctld or the system generation is replaced. They gain the new tracing namespace and container-local /sys/fs/bpf mount on their next normal start, with a kernel that provides tracing namespaces. Activation does not retrofit those namespaces or require stopping workloads to obtain them. The tracing namespace is omitted when the running kernel does not provide it. The host's private bpffs is not the writable filesystem exposed inside the container.

Userspace downgrade considerations

Before downgrading to osctld without NetworkManager resolver ownership, unset managed resolvers with the newer daemon while the guest is running. Verify that guest DNS has resumed its normal policy. The older daemon cannot remove the newer daemon's owned 10-osctl-dns.conf drop-in. Preserve operator-owned NetworkManager configuration rather than deleting all drop-ins.

Per-container bpffs mounts created by the newer daemon are unknown to older userspace and can remain after a downgrade until the next host reboot. This is a bounded rollback resource leak, not a reason to drain containers during forward userspace activation. Do not remove a live container's mounts merely to downgrade management userspace.

Container syslog namespace tags are normalized to the kernel ABI and include a run-specific suffix. Host log processing must not assume that the old container/pool label is unchanged between normal container starts.

Kernel livepatch coverage

Livepatch availability and vulnerability coverage are specific to the exact boot kernel. Follow the kernel livepatch release process and the target kernel's release page before activation.

The former Linux 6.12.95 livepatch line is no longer built or managed by current configurations. Removing its source support does not unload a module that is already active. Keep its raw sysfs state visible, fail over or fence the host when required, and reboot into approved native code. Do not force-disable or remove the module. Retained system generations and store closures remain available for recovery.