Composable VM Images for Kata Containers

Status: Implemented

Summary

This proposal introduces a composable image architecture for Kata Containers guest VMs. Instead of building a single monolithic rootfs image that contains every component a workload might need, the runtime assembles a VM from a lean base image plus zero or more purpose-specific guest extension images that are cold-plugged as additional virtio-blk devices. Each extension is an EROFS image protected by dm-verity, mounted read-only inside the guest before the kata-agent starts.

The first application of this architecture splits Confidential Containers (CoCo) guest components out of the monolithic kata-containers-confidential.img into a separate kata-containers-coco-extension.img.

The current proposal targets QEMU as the hypervisor backend. The design is intentionally hypervisor-agnostic at the configuration and guest layers — the guest_extension_images configuration, guest-side systemd units, and agent path resolution work identically regardless of the hypervisor. Only the host-side device attachment is hypervisor-specific (QEMU virtio-blk cold-plug today), and extending support to other hypervisor backends (Cloud Hypervisor, Firecracker, etc.) requires only implementing the equivalent block device attachment in each backend.

Motivation

Today every CoCo runtime class ships a self-contained rootfs image that bundles the base OS, the kata-agent, and every CoCo-specific binary (attestation-agent, confidential-data-hub, api-server-rest, ocicrypt config, pause bundle). This creates several problems:

1. Image bloat — the confidential image is significantly larger than the standard image because it carries components that are only needed for confidential workloads.

2. Combinatorial explosion — adding new feature dimensions (e.g. GPU support, different TEE backends) multiplies the number of monolithic images that must be built, tested, and distributed.

3. Slow iteration — updating a single CoCo binary requires rebuilding and re-signing the entire rootfs image.

4. Lack of composability — users who need a custom combination of features must maintain their own image build pipeline.

A composable approach addresses all four issues: the base image stays small and generic, guest extension images are independently built and versioned, and the runtime composes them at VM creation time.

Design

Architecture overview

Host                                          Guest VM
┌───────────────────────────────┐
│  Runtime config               │
│  ┌─────────────────────────┐  │
│  │ image = base            │  │      ┌─────────────────────────────────────┐
│  │ guest_extension_images: │──┼─────>│  /  (base rootfs, erofs)            │
│  │   - name: coco          │  │ boot │    kata-agent, systemd, ...         │
│  │     path: ...           │  │ cold │                                     │
│  │     verity: ...         │  │ plug │  /run/kata-extensions/coco/ (erofs) │
│  └─────────────────────────┘  │      │    attestation-agent                │
│                               │      │    confidential-data-hub            │
│  QEMU                         │      │    api-server-rest                  │
│    -drive base.img            │      │    ocicrypt_config.json             │
│    -drive coco.img            │      │    pause_bundle/                    │
│    serial=extension-coco      │      │                                     │
└───────────────────────────────┘      │  dm-verity protects both the        │
                                       │  base rootfs and the extension.     │
                                       └─────────────────────────────────────┘

Runtime configuration

A new guest_extension_images field on the hypervisor configuration accepts an ordered list of guest extension images:

[hypervisor.qemu]
image = "/opt/kata/share/kata-containers/kata-containers.img"

# ...existing keys...

[[hypervisor.qemu.guest_extension_images]]
name = "coco"
path = "/opt/kata/share/kata-containers/kata-containers-coco-extension.img"
verity_params = "root_hash=abc...,salt=def...,data_blocks=1234,hash_block_size=4096,data_block_size=4096"

Each entry maps to a Rust struct:

pub struct GuestExtensionImage {
    pub name: String,
    pub path: String,
    pub verity_params: String,
}

The name field is the primary identifier for an extension. It must be unique within the configuration and is used to:

• Set the virtio-blk serial to extension-<name>, enabling deterministic device discovery in the guest via /sys/block/*/serial.
• Namespace the kernel command-line verity parameters as kata.extension.<name>.verity_params=....
• Name the guest mount point at /run/kata-extensions/<name>/.
• Name the dm-verity device mapper target as extension-<name>.

Host-side: cold-plugging extension devices

At VM creation time, for each entry in guest_extension_images, the runtime:

1. Creates a virtio-blk device backed by the guest extension image file. Extensions are always attached as virtio-blk, using the architecture's virtio-blk transport (virtio-blk-ccw on s390x, virtio-blk-pci elsewhere). Neither the VM rootfs driver (vm_rootfs_driver) nor the generic block device driver (block_device_driver) is reused here: those may resolve to a non-virtio-blk transport such as virtio-pmem (NVDIMM) or virtio-scsi, and only virtio-blk devices carry the serial the guest relies on for discovery (step 2), so a non-virtio-blk transport would leave the extension undiscoverable and its mount unit would fail closed.
2. Sets the device serial to extension-<name>.
3. Appends kata.extension.<name>.verity_params=<value> to the guest kernel command line for every configured extension. The value is empty for an unmeasured extension (verity_params = "", e.g. on s390x — see "Mount script" below); the entry is emitted unconditionally because it doubles as the guest-side activation signal for the mount unit. With an empty value the entry renders as a bare kata.extension.<name>.verity_params (no =), which the unit condition and generator both handle.

Both the Go runtime and the Rust runtime-rs implement this logic in their respective QEMU hypervisor backends. The extension devices are cold-plugged — they are present on the QEMU command line at VM start, not hot-plugged later.

Other hypervisor backends (Cloud Hypervisor, Firecracker) do not implement extension cold-plugging yet, but the mechanism is straightforward to add: each backend only needs to translate the GuestExtensionImage entries into its native block device attachment API. The runtime configuration, guest-side mounting, and agent path resolution are completely hypervisor-independent.

Guest-side: mounting extensions

A systemd template unit [email protected] handles extension discovery and mounting. The unit is instantiated per extension name (e.g. [email protected]); a systemd generator (described below) creates those instances automatically from the kernel command line.

The mounting is performed by a systemd unit (before the agent starts) rather than by the kata-agent itself for several reasons:

• Chicken-and-egg problem — in the CoCo extension case, the extension carries the very guest components (attestation-agent, confidential-data-hub) that the kata-agent needs to launch. The agent cannot mount the extension because it needs the extension's contents to be available before it starts.

• Init-system ordering guarantees — systemd provides declarative ordering (Before=, After=), conditional activation (ConditionKernelCommandLine=), and failure handling (OnFailure=poweroff.target). Reimplementing these guarantees inside the agent would duplicate init-system responsibilities.

• Separation of concerns — block device discovery, dm-verity setup, and filesystem mounting are OS-level operations that belong in the init layer. The agent's role is to consume the mounted filesystem, not to manage block devices.

• Non-systemd environments — the design does not mandate systemd. In environments using a different init system (or a dedicated orchestrator like NVRC), the equivalent mounting logic can be implemented by whatever entity manages early boot. The key requirement is that guest extension images are mounted before the kata-agent starts — how that is achieved is an init-system concern.

Service unit

[Unit]
Description=Mount Kata guest extension image %i
DefaultDependencies=no
Before=kata-agent.service
After=local-fs-pre.target
ConditionKernelCommandLine=kata.extension.%i.verity_params
OnFailure=poweroff.target

[Service]
Type=oneshot
RemainAfterExit=yes
ExecStart=/usr/libexec/kata-extension-mount.sh %i
ExecStop=/usr/libexec/kata-extension-umount.sh %i

[Install]
WantedBy=kata-containers.target

Key design decisions:

• ConditionKernelCommandLine — the service only activates when the runtime has actually configured the extension. This prevents the unit from running (and failing) in non-extension VM configurations.

• Before=kata-agent.service — guarantees the extension filesystem is mounted before the agent attempts to use any component from it.

• OnFailure=poweroff.target — if the extension cannot be mounted (e.g. verity verification failure, missing device), the VM is shut down immediately. A confidential VM must not continue with an unverified or missing extension.

Enabling instances

Implementation note. An earlier revision enabled a fixed [email protected] symlink at rootfs build time, which would have required editing the image build for every new extension. Review feedback pushed us to the generator below so the build stays extension-agnostic.

systemd template units must be enabled per instance, and the set of extensions is only known at runtime (from the kernel command line). A systemd generator, kata-extension-mount-generator, bridges that gap: it runs in early boot, scans /proc/cmdline for every kata.extension.<name>.verity_params entry (whether it carries a value or, for an unmeasured extension, is a bare key), and symlinks kata-extension-mount@<name>.service into kata-containers.target.wants.

Because the runtime emits exactly one such cmdline entry per configured guest_extension_images, the generator enables precisely the extensions the VM was configured with. Adding a new extension therefore requires no change to the rootfs build — it is wired up entirely from runtime configuration. The generator ships in the base image (installed alongside the template unit by the agent), so it is part of the dm-verity-measured rootfs, and the cmdline it reads is itself covered by the guest launch measurement.

Mount script

The mount script (kata-extension-mount.sh) performs the following steps:

1. Device discovery — scans /sys/block/*/serial for a device whose serial matches extension-<name>. This is more reliable than waiting for udev to create /dev/disk/by-id/ symlinks, since the minimal guest environment may not run a full udev daemon.

2. Verity parameter extraction — reads kata.extension.<name>.verity_params=... from /proc/cmdline and parses the comma-separated key=value pairs (root_hash, salt, data_blocks, hash_block_size, data_block_size).

3. dm-verity setup — runs veritysetup open with --no-superblock to open the verity target. The --no-superblock flag is required because the image builder uses veritysetup format --no-superblock during image creation — all verity parameters are passed explicitly rather than stored in an on-disk superblock.

4. EROFS mount — mounts the resulting device read-only at /run/kata-extensions/<name>/ (the dm-verity target when verified, or the data partition directly for an unmeasured extension).

Integrity policy: measured vs. unmeasured, and failing closed

An extension can legitimately ship without dm-verity — for example on s390x, where IBM Secure Execution protects the guest through a different mechanism and images are built with MEASURED_ROOTFS=no. The mount script must therefore support a raw (unverified) mount, but it must not let that path become a silent downgrade: an attacker who can edit the (host-supplied) kernel command line could otherwise strip verity_params from a measured extension and have it mounted unverified.

The script separates these two cases explicitly, using the on-disk layout as the source of truth rather than trusting the cmdline alone. The image build encodes its integrity policy in the partition table (see "Image build" below): a measured extension carries a dm-verity hash partition (p2) next to the data partition (p1), while an unmeasured extension has only p1. The script detects the presence of the hash device and cross-checks it against the verity_params carried on the kernel command line (which, in a confidential guest, is itself part of the measured, attested boot):

Any "refuse" path exits non-zero; combined with OnFailure=poweroff.target on the mount unit, that powers the VM off rather than continuing with an unverified or inconsistent extension. The defence has two layers: the cmdline (and thus the root hash, or its deliberate absence) is covered by the guest launch measurement and remote attestation, and the in-guest layout cross-check fails closed so a measured extension can never be silently downgraded to a raw mount.

Agent-side: data-driven component manifest

Implementation note. An earlier revision of this proposal described the agent as resolving a small, hardcoded list of component paths. While implementing and testing the extension we found this too rigid: the extension evolved to ship multiple attestation-agent flavours, per-process environment requirements, and ordering constraints that the agent should not have to know about. The design below — a data-driven manifest owned by the extension — is what we converged on so that adding or reconfiguring a bundled component requires no kata-agent code change.

Each extension ships a manifest at etc/kata-extensions/components.toml. When the extension is mounted, the kata-agent reads it and builds its launch plan from it. The manifest declares the processes to launch and the resources they expose; all paths are relative to the extension mount root (/run/kata-extensions/<name>).

A process entry carries:

• id and a level used to order and gate launches.
• args (and optional_args, which are appended only when a named context variable is non-empty).
• config, wait_socket (the agent blocks until the socket appears), and timeout_secs.
• env — extra environment variables for the spawned process.
• An optional select selector plus [process.variants.<name>] tables, so a single logical component can ship multiple binaries and the consumer picks one at runtime.

${...} tokens in args, config, env values, and variant fields are substituted by the agent from a runtime context it assembles (socket and config paths, the extension mount root ${extension_root}, the selected ${attester_variant}, etc.). Introducing a brand-new variable is the only change that ever needs to touch the agent.

Abridged manifest (see "Attester variant selection" and "Runtime dependencies" below for why the nvidia variant and the PATH/LD_LIBRARY_PATH entries are shaped the way they are):

schema_version = 1

[paths]
"ocicrypt-config" = "etc/ocicrypt_config.json"
"pause-bundle"    = "pause_bundle"

[[process]]
id          = "attestation-agent"
level       = 1
args        = ["--attestation_sock", "${aa_attestation_uri}"]
config      = "${aa_config_path}"
wait_socket = "${aa_attestation_socket}"
select      = "${attester_variant}"

  [process.variants.default]
  path = "usr/local/bin/attestation-agent"

  [process.variants.nvidia]
  path = "usr/local/bin/attestation-agent-nv"
  env  = { LD_LIBRARY_PATH = "${extension_root}/usr/local/lib:/run/kata-extensions/gpu/usr/lib" }

[[process]]
id          = "confidential-data-hub"
level       = 2
path        = "usr/local/bin/confidential-data-hub"
config      = "${cdh_config_path}"
env         = { OCICRYPT_KEYPROVIDER_CONFIG = "${ocicrypt_config_path}", PATH = "${extension_root}/usr/sbin:/bin:/sbin:/usr/bin:/usr/sbin" }
wait_socket = "${cdh_socket}"

[[process]]
id   = "api-server-rest"
level = 3
path = "usr/local/bin/api-server-rest"
args = ["--features", "${rest_api_features}"]

When no extension is mounted, the agent falls back to a built-in launch plan that reproduces the legacy behaviour (components launched from /usr/local/bin/... in the rootfs). The same dual-path principle applies to the non-process resources declared under [paths]: the agent resolves them inside the extension first and falls back to the legacy location otherwise:

This approach:

• Preserves backward compatibility with existing monolithic rootfs images where CoCo components are baked into the base image.
• Requires no special rootfs modifications — the base image does not need stub files or directories for the extension components.
• Works transparently on a read-only rootfs — no bind-mounting, no remounting, no writes to the root filesystem.
• Keeps the agent generic — extension-specific names, env, and ordering live in the manifest, not in agent code.

Attester variant selection and the NVRC contract

The CoCo extension ships two attestation-agent builds: the stock attestation-agent and an NVIDIA-attester build, attestation-agent-nv, that collects GPU evidence in addition to the TEE evidence. Which one runs is chosen by the manifest's select = "${attester_variant}" selector, and the value of ${attester_variant} is driven by the guest init:

• On a plain confidential guest the kata-agent runs init itself and the variable defaults to default → the stock attester launches.
• On a GPU guest, NVRC (the NVIDIA runtime config that owns early boot) detects the GPU extension and exports KATA_ATTESTER_VARIANT=nvidia before exec'ing the kata-agent. The agent forwards that into the substitution context as ${attester_variant}, so the nvidia variant launches.

This is a small cross-component contract: the environment variable name (KATA_ATTESTER_VARIANT) and the nvidia value are produced by NVRC and consumed by the kata-agent, which keeps the agent free of any GPU- or NVIDIA-specific knowledge — it only knows how to forward a selector into the manifest. We arrived at this split after first trying to special-case the attester inside the agent; pushing the decision out to NVRC + the manifest kept both the agent and the extension generic.

Why one extension with two builds (and not two extensions)

Shipping two attestation-agent builds inside a single CoCo extension is a deliberate choice, and it is worth being explicit about it because extensions are otherwise meant to eliminate duplication.

• A single CoCo extension serves both plain confidential guests (TEE evidence only) and confidential GPU guests (TEE + GPU evidence). The only thing that differs between them is which attestation-agent binary runs; everything else in the extension (confidential-data-hub, api-server-rest, ocicrypt config, pause bundle, cryptsetup) is shared verbatim. Splitting the NVIDIA attester into its own extension would duplicate that shared payload and force every confidential GPU guest to compose two CoCo-flavoured extensions instead of one.
• The cost of keeping both builds together is precisely the manifest's select/variants machinery: the manifest has to be aware that the attestation-agent comes in two flavours and pick one at runtime. We consider that a fair trade — the complexity is confined to data (the manifest), the agent stays generic, and the extension stays a single, coherent "CoCo" unit.
• A separate extension only pays off when its contents are substantially different (e.g. the GPU extension, which carries the driver userspace), not for two near-identical builds of the same component. No further attester variants are planned today, but if one appeared it would be another [process.variants.<name>] entry — not a new image.

Runtime dependencies: dynamic linking and secure-mount tooling

Two classes of runtime dependency surfaced only once the components actually ran inside a composed VM. Both are resolved by the manifest env entries above rather than by agent code, and both informed where binaries and libraries must physically live.

NVIDIA attester dynamic libraries

attestation-agent-nv links libnvat.so (the NVIDIA Attestation SDK), which in turn:

• dlopens libnvidia-ml.so.1 (NVML) at runtime to gather GPU evidence. NVML is part of the GPU extension, mounted at /run/kata-extensions/gpu with its driver libraries under usr/lib.
• pulls in a closure of non-glibc libraries (libxml2, zlib, lzma, the C++ runtime, ...) that the guest rootfs does not otherwise ship.

Neither set is present in a stock guest, so:

• The CoCo extension build bundles libnvat.so and every non-glibc transitive dependency next to it under usr/local/lib.
• The nvidia manifest variant (the [process.variants.nvidia] entry, not a separate image) sets an LD_LIBRARY_PATH that lists both the CoCo extension's usr/local/lib (for libnvat and its closure) and the GPU extension's usr/lib (for NVML). Setting only the first was the cause of an NVAT Error 500: NVML Initialization Failed we hit during bring-up; the RCAR handshake then never produced GPU evidence and attestation failed.

Because the agent applies manifest env on top of the inherited environment (without clearing it), and because the nvidia variant only ever runs when the GPU extension is present, referencing the GPU extension's well-known mount path here is safe.

CDH secure-mount tooling (encrypted vs plain storage)

The confidential-data-hub secure_mount feature shells out — by PATH lookup — to external tools, and these split cleanly into two buckets that belong in two different places:

• Plain storage setup — mke2fs/mkfs.ext4 (and dd, plus /etc/mke2fs.conf) format the scratch volume. This is needed for unencrypted ephemeral storage too, so it belongs in the base image and ships there unconditionally.
• Encrypted storage — cryptsetup LUKS-formats and opens the volume. This is a CoCo-only capability, so it belongs with the CoCo guest components in the coco extension.

This is the same veritysetup-vs-cryptsetup reasoning already applied to extension mounting: the base must always carry veritysetup (it opens every extension as a dm-verity device before mounting), and cryptsetup shares an identical shared-library closure, so wherever veritysetup lives the libraries for cryptsetup are already present.

Because CDH runs in the base rootfs namespace but cryptsetup lives in the extension (which is not on the default search path), the manifest sets CDH's PATH to ${extension_root}/usr/sbin:/bin:/sbin:/usr/bin:/usr/sbin — the extension's cryptsetup first, then the base directories that carry mke2fs/mkfs.ext4/dd. (The kata-agent launches components with PATH=/bin:/sbin:/usr/bin:/usr/sbin; since setting any env value replaces that variable wholesale, the base directories are restored explicitly.)

How each tool is provisioned depends on the base flavour (see "Base image flavours" below), but the placement contract is identical:

The distroless path needs explicit copying because nothing lands in a chiseled image unless placed there deliberately, and the NVIDIA base is never built with CONFIDENTIAL_GUEST=yes. The full-distro base is built with CONFIDENTIAL_GUEST=yes, so the same tools arrive as ordinary packages. In both cases the extension's cryptsetup resolves its libraries against the base, which requires the base and the coco-extension builder to stay on the same distro/ABI (Ubuntu 24.04 "noble" today).

Image build pipeline

Base image flavours

Kata base images come in two flavours, distinguished by who owns early boot. This distinction — not "Ubuntu vs NVIDIA" — is what drives the differences in how tooling is provisioned and who mounts the extensions:

• Full-distro base — ships a complete distribution with systemd as init. systemd discovers and mounts the extensions (via [email protected]), and the binaries/libraries the guest needs arrive as ordinary distribution packages. The standard confidential kata-containers.img (Ubuntu) is today's instance.
• Distroless base — a minimal, chiseled image with no full init system. A dedicated early-boot component takes over the responsibilities systemd would otherwise have (extension discovery and mounting, attester selection, orchestration), and any tooling must be placed into the image deliberately rather than pulled in as packages. The chiseled nvidia base driven by NVRC is today's instance.

These are the only two flavours today and no others are planned, but a new base would fall into one of these categories and follow the same mechanisms (systemd units for a full-distro base; an NVRC-like early-boot owner for a distroless one). The sections below describe the build for both; where they diverge it is because of the flavour, not because the NVIDIA image is a special case.

Base image

The full-distro kata-containers.img is built as before, but without CoCo guest components (attestation-agent, confidential-data-hub, api-server-rest, ocicrypt config, pause bundle). It includes:

• The kata-agent
• systemd and the [email protected] template unit
• cryptsetup-bin (provides veritysetup) — required unconditionally so that the base image can mount verity-protected extensions regardless of whether the base itself was built with CONFIDENTIAL_GUEST=yes. On Ubuntu this same package also provides cryptsetup, so the full-distro base happens to carry the encrypted-storage binary too.
• The plain-storage tooling for CDH secure_mount — mke2fs/mkfs.ext4, dd, and /etc/mke2fs.conf. On the standard base these come from e2fsprogs (added when CONFIDENTIAL_GUEST=yes) and coreutils. See "Runtime dependencies" for why the encrypted-storage cryptsetup lives in the extension instead.

The distroless base (nvidia, driven by NVRC) is a chiseled, driver-agnostic image rather than a full distro, so the items above do not arrive as packages — they are copied into the base layout explicitly: veritysetup and its library closure unconditionally, and the mke2fs/mkfs.ext4/dd/mke2fs.conf plain-storage tooling alongside it.

CoCo guest extension image

The kata-containers-coco-extension.img is built by:

1. Unpacking the CoCo guest components tarball into a temporary rootfs. Besides the agent-launched binaries (attestation-agent, attestation-agent-nv, confidential-data-hub, api-server-rest) this tarball also carries:
   • cryptsetup under usr/sbin — the encrypted-storage binary for CDH secure_mount (its shared libraries are resolved against the base; see "Runtime dependencies").
   • the NVIDIA attester libraries under usr/local/lib — libnvat.so plus its non-glibc transitive closure (libxml2, zlib, lzma, the C++ runtime).
2. Unpacking the pause image tarball into the same rootfs.
3. Writing the component manifest to etc/kata-extensions/components.toml.
4. Running the image builder with:
   • FS_TYPE=erofs — EROFS filesystem for compact, read-only storage.
   • MEASURED_ROOTFS=yes — creates a dm-verity hash partition.
   • SKIP_DAX_HEADER=yes — no DAX header (virtio-blk, not NVDIMM).
   • SKIP_ROOTFS_CHECK=yes — the extension has no /sbin/init.
   • BUILD_VARIANT=coco-extension — produces a correctly named root hash file (root_hash_coco-extension.txt).

The resulting image is a two-partition disk:

• Partition 1: EROFS data partition containing the CoCo components.
• Partition 2: dm-verity hash partition.

The root hash and verity parameters are captured at build time and injected into the runtime configuration templates.

Security model

The extension architecture preserves the existing security guarantees of Confidential Containers:

• dm-verity provides integrity protection for the guest extension image, identical to the existing protection on the base rootfs. Any tampering with the extension contents is detected at mount time.

• Verity parameters on the kernel command line are measured by the TEE firmware (OVMF/TDVF for TDX, SEV-SNP firmware) as part of the launch measurement. An attacker cannot substitute different verity parameters without changing the measurement, which would be detected during attestation.

• Mount failure = VM shutdown — OnFailure=poweroff.target ensures the VM does not proceed with missing or tampered components.

• Read-only mounts — both the base rootfs and guest extension images are mounted read-only (EROFS), preventing runtime modification.

Alternatives considered

systemd-sysext

systemd provides a built-in mechanism for composable system extensions via systemd-sysext. System extension images are overlaid onto /usr/ and /opt/ using overlayfs, making their contents appear as if they were part of the base OS image. The mechanism supports EROFS, squashfs, and ext4 images, and can verify them with dm-verity.

sysext was evaluated and is conceptually aligned with this proposal. However, several gaps make it unsuitable as a direct replacement for the custom mounting approach described here:

• Hierarchy coverage — sysext only merges into /usr/ and /opt/. Files under /etc/ (e.g. ocicrypt_config.json) require the separate confext mechanism, and files outside these three hierarchies (e.g. /pause_bundle/) cannot be delivered via sysext or confext at all.

• Block device bridging — sysext discovers images from well-known directories (/var/lib/extensions/, /run/extensions/), not from raw block devices. A bridging step would still be needed to discover the virtio-blk extension device, verify it with dm-verity, and place or symlink it where sysext expects to find it — eliminating much of the simplification sysext would offer.

• Boot ordering — systemd-sysext.service merges extensions before basic.target, but the precise ordering relative to kata-agent.service is not directly controllable. The current Before=kata-agent.service guarantee on the custom mount unit gives explicit control over this.

• Version matching — sysext enforces extension-release.d metadata matching against the host os-release. While this is useful for general purpose systems, it adds friction in the Kata context where the base image and guest extension images are built and versioned together in a controlled pipeline.

• Minimal guest environment — the Kata guest rootfs is a minimal environment that may not ship a full systemd with sysext/confext support enabled.

The current proposal can be evolved toward sysext in the future if these gaps are addressed, particularly if sysext gains support for block-device-backed extensions or if the Kata guest components are restructured to fit entirely within /usr/.

bootc and ostree

bootc and ostree provide image-based OS deployment and update mechanisms. bootc in particular enables building bootable container images and managing OS updates as container image pulls.

While bootc and ostree share the high-level goal of composable, image-based OS management, they solve a fundamentally different problem than this proposal:

• bootc/ostree manage the base OS lifecycle — they address how the root filesystem is built, deployed, and updated over time (e.g. atomic upgrades, rollbacks, image pulls).

• This proposal manages VM-time composition — it addresses how purpose-specific components are attached to an already-built base image at VM creation time, without modifying the base image itself.

The two approaches are orthogonal. A Kata base image could be built and managed using bootc/ostree, and guest extension images would still be cold-plugged and mounted at VM boot using the mechanism described here. bootc does not provide a mechanism for dynamically composing additional block devices into a running or booting system — it operates at the image build and deployment layer, not at the VM assembly layer.

Future work

Additional extension types

The architecture is designed to support multiple guest extension images:

• GPU extension — the NVIDIA GPU userspace (driver libraries, NVML, the container-toolkit binaries, kernel modules) lives in a gpu-extension image mounted at /run/kata-extensions/gpu, carved out of the same build as the driver-agnostic nvidia base image. NVRC orchestrates early boot, loads the modules from the extension, and composes the GPU extension with the CoCo extension on confidential GPU guests. This extension is implemented; its interplay with the CoCo extension (NVML resolution, attester selection) is covered in "Runtime dependencies" and "Attester variant selection" above.

• Custom extensions — users can build their own guest extension images for workload-specific libraries, models, or configurations.

Additive image assembly

Today the NVIDIA images are carved out of a single chiseled monolith tree: the gpu-extension is assembled additively (an allow-list of GPU userspace is copied into a fresh tree), while the nvidia base is produced subtractively — the same allow-list is deleted from the full tree. This keeps the monolith byte-identical during the transition, but it means the base build describes what it does not want rather than what it does.

The intended next step is to invert this into a purely additive flow, so nothing is ever subtracted:

1. Build the shared driver stage-one (always required).
2. Assemble the nvidia base additively from stage-one.
3. Assemble the gpu-extension additively from stage-one.
4. Compose the monolith by combining the base and the extension.

This requires splitting the interleaved chisseled_* producers (which today copy base runtime libraries and GPU userspace in the same pass) into base and GPU halves. It is best done together with the removal of the Go-runtime monolith, so that only a single, additive code path remains rather than maintaining both the monolith and the split builds.

Extension ordering

When multiple extensions are configured, they are cold-plugged in the order they appear in the guest_extension_images list. The current design does not enforce explicit ordering dependencies between extensions. If future use cases require extension-to-extension dependencies, the systemd units can be extended with appropriate After=/Requires= relationships.

Other hypervisor backends

The current proposal covers QEMU only. Extending to other backends requires implementing block device cold-plug for each:

• Cloud Hypervisor — add --disk entries with the guest extension image path and serial. Cloud Hypervisor natively supports virtio-blk serial numbers.
• Dragonball — attach additional virtio-blk devices through the Dragonball VMM's block device configuration, mapping each GuestExtensionImage to a drive with the corresponding serial.
• Firecracker — add block device entries via the Firecracker API with the appropriate drive ID. Serial-based discovery may need adaptation since Firecracker exposes drive IDs differently.

No changes are needed in the guest-side mounting logic or the agent — the extension device discovery via /sys/block/*/serial and the systemd units work the same way regardless of which hypervisor attached the block device.

Manifest-driven extension discovery via init-data

The current design passes verity parameters for each extension on the kernel command line. This works well for a small number of extensions but does not scale: each additional extension adds a long kata.extension.<name>.verity_params=... entry, and the kernel command line has practical size limits.

A future evolution could introduce a two-phase bootstrap:

1. Phase 1 (kernel-params driven) — the kernel command line carries verity parameters for a single, small manifest extension image. This image is mounted first using the existing mechanism.

2. Phase 2 (manifest driven) — the manifest extension contains a configuration file (e.g. extensions.conf) listing all other extensions with their names, verity parameters, and any additional metadata. The mount script reads this file and mounts the remaining extensions accordingly.

This approach has several advantages:

• Kernel command line stays fixed-size regardless of how many extensions are composed.
• Attestation is simplified — the TEE firmware measures one manifest hash on the command line; the verifier only needs to validate that single hash. The chain of trust extends from the measured kernel command line to the verified manifest to the verified extensions.
• Richer metadata — the manifest can carry structured information beyond verity parameters: extension ordering constraints, version requirements, policies, or init-data payloads.

The design of the systemd units and mount script can support both modes: if per-extension verity parameters are present on the kernel command line, they are used directly (current behavior); if a manifest extension is present, it is consulted for the remaining extensions. This makes the manifest mode an additive, backward-compatible evolution.

Extension versioning and attestation

Future work may add version metadata to guest extension images, enabling the attestation flow to verify not just integrity (via dm-verity) but also that the specific expected version of each component is present.

References

• Kata Containers architecture
• dm-verity documentation
• EROFS filesystem
• systemd-sysext
• bootc
• ostree