Skip to navigation

NVIDIA Workload Identity Controller

View as Markdown

dsx-identity-controller watches DPU, Launch Layer VirtualInstance, and Launch Layer BareMetalInstance objects in one management cluster. It derives the corresponding SPIRE ClusterStaticEntry objects without modifying its sources.

The Identity Controller owns non-Pod registration fields. The official SPIRE controller-manager owns Pod registrations. The controller rejects duplicate child SPIFFE IDs and duplicate VM Pod UIDs, removes stale entries for invalid sources, and reconciles again when the source or conflicting entry changes.

Every DPU source identifies its paired operator-owned BM through the carbide.nvidia.com/dpu-machine-id label. A complete tenant BareMetalInstance for that DPU replaces the operator fallback entry. The launchlayer.nvidia.com/org-id annotation supplies the tenant.

A VirtualInstance supplies the VM instance ID, worker node name, and current virt-launcher Pod UID. Its identity is parented to the operator-owned worker’s dsx_imds agent and selected only by that Pod UID.

The official controller-manager registers each Kubernetes Pod beneath its worker’s dsx_imds-attested agent. Its registration uses the configured operator tenant, Kubernetes node name, and Pod UID, and the Kubernetes WorkloadAttestor serves it. The paired BM registration is parented to the external dpu_hw-attested agent identified by the DPU serial number.

SPIRE registration API DPU custom resource NVIDIA Workload IdentityController Deployment VirtualInstance BareMetalInstance SPIRE ClusterStaticEntry custom resource active SPIRE controller-manager replica Shard-local SPIRE server

The controller runs in the regular management cluster that owns the identity shard’s source CRs through charts/dsx-identity/templates/identity-controller.yaml.

Flags

FlagDefaultPurpose
--trust-domainrequiredLocal SPIRE trust domain
--operator-tenantrequiredConfigured tenant for operator identities
--class-namerequiredSPIRE controller-manager class name
--dpu-metadata-selectorrequired, repeatableKubernetes WorkloadAttestor selector for DPU IMDS
--source-api-poll-interval1hPositive duration between discovery retries for absent source APIs
--metrics-bind-address:8080Controller metrics address
--health-probe-bind-address:8081/healthz and /readyz address
--leader-electtrueEnable Kubernetes leader election

The command uses the active Kubernetes client configuration and requires the SPIRE CRDs. DPU and Launch Layer source watches start when their APIs exist. The controller checks APIs at startup and retries absent APIs every hour by default. Set --source-api-poll-interval to change the retry period. It registers watches when the CRDs are installed. Existing source objects reconcile when the watch starts. Without the BareMetalInstance API, DPU sources retain their operator-owned BM identities. Its assignment watch replaces those identities when complete tenant assignments appear. Discovery and access errors remain errors.