Add nvidia DRA driver GPU service

Release.toml

@@ -477,5 +477,6 @@ version = "1.64.0"
 ]
 "(1.63.0, 1.64.0)" = [
     "migrate_v1.64.0_kubernetes-container-runtime-endpoint.lz4",
-    "migrate_v1.64.0_hugepages-settings.lz4"
+    "migrate_v1.64.0_hugepages-settings.lz4",
+    "migrate_v1.64.0_kubelet-dra-drivers-nvidia-enabled.lz4"
 ]

sources/Cargo.toml

@@ -67,6 +67,7 @@ members = [
     "settings-migrations/v1.63.0/nvidia-k8s-device-plugin-enabled",
     "settings-migrations/v1.64.0/kubernetes-container-runtime-endpoint",
     "settings-migrations/v1.64.0/hugepages-settings",
+    "settings-migrations/v1.64.0/kubelet-dra-drivers-nvidia-enabled",
 
     "settings-plugins/aws-dev",
     "settings-plugins/aws-ecs-2",

sources/settings-defaults/aws-k8s-1.33-nvidia/defaults.d/83-nvidia-dra-driver-gpu.toml

@@ -0,0 +1 @@
+../../../shared-defaults/nvidia-dra-driver-gpu.toml

sources/settings-defaults/aws-k8s-1.36-nvidia/defaults.d/83-nvidia-dra-driver-gpu.toml

@@ -0,0 +1 @@
+../../../shared-defaults/nvidia-dra-driver-gpu.toml

sources/settings-migrations/v1.64.0/kubelet-dra-drivers-nvidia-enabled/Cargo.toml

@@ -0,0 +1,10 @@
+[package]
+name = "kubelet-dra-drivers-nvidia-enabled"
+version = "0.1.0"
+license = "Apache-2.0 OR MIT"
+edition = "2021"
+publish = false
+exclude = ["README.md"]
+
+[dependencies]
+migration-helpers.workspace = true

sources/settings-migrations/v1.64.0/kubelet-dra-drivers-nvidia-enabled/src/main.rs

@@ -0,0 +1,27 @@
+use migration_helpers::common_migrations::AddSettingsMigration;
+use migration_helpers::{migrate, Result};
+use std::process;
+
+/// We added a new setting for enabling/disabling the NVIDIA DRA driver
+/// (gpu-kubelet-plugin) host service:
+/// - settings.kubelet-dra-drivers.nvidia.enabled
+///
+/// Forward (upgrade): no-op. The `enabled` field defaults to `None` (the
+/// service treats unset as disabled), so existing data is compatible.
+/// Backward (downgrade): removes settings.kubelet-dra-drivers.nvidia.enabled from the
+/// datastore so older versions don't encounter an unknown field.
+fn run() -> Result<()> {
+    migrate(AddSettingsMigration(&[
+        "settings.kubelet-dra-drivers.nvidia.enabled",
+    ]))
+}
+
+// Returning a Result from main makes it print a Debug representation of the error, but with Snafu
+// we have nice Display representations of the error, so we wrap "main" (run) and print any error.
+// https://github.com/shepmaster/snafu/issues/110
+fn main() {
+    if let Err(e) = run() {
+        eprintln!("{e}");
+        process::exit(1);
+    }
+}

sources/settings-plugins/aws-k8s-nvidia/src/lib.rs

@@ -27,4 +27,5 @@ struct AwsK8sSettings {
     autoscaling: bottlerocket_settings_models::AutoScalingSettingsV1,
     nvidia_container_runtime: bottlerocket_settings_models::NvidiaContainerRuntimeSettingsV1,
     kubelet_device_plugins: bottlerocket_settings_models::KubeletDevicePluginsV1,
+    kubelet_dra_drivers: bottlerocket_settings_models::KubeletDraDriversV1,
 }

sources/shared-defaults/nvidia-dra-driver-gpu.toml

@@ -0,0 +1,27 @@
+# NVIDIA DRA driver (GPU kubelet plugin) service.
+#
+# The plugin runs as a host systemd service (not a Helm-installed DaemonSet).
+# Its ExecStart is rendered from settings because it needs the node name from
+# settings.kubernetes.hostname-override; enablement is the dedicated
+# settings.kubelet-dra-drivers.nvidia.enabled toggle (independent of the classic
+# device plugin's settings.kubelet-device-plugins.nvidia.enabled).
+[services.nvidia-dra-driver-gpu]
+restart-commands = [
+    "/bin/systemctl try-reload-or-restart nvidia-dra-driver-gpu.service",
+]
+configuration-files = [
+    "nvidia-dra-driver-gpu-exec-start-conf",
+]
+
+[configuration-files.nvidia-dra-driver-gpu-exec-start-conf]
+path = "/etc/systemd/system/nvidia-dra-driver-gpu.service.d/exec-start.conf"
+template-path = "/usr/share/templates/nvidia-dra-driver-gpu-exec-start-conf"
+
+# Opt-in by default: the DRA driver is off unless explicitly enabled, so it does
+# not conflict with the classic device plugin out of the box.
+[settings.kubelet-dra-drivers.nvidia]
+enabled = false
+
+# Re-render + restart the service when the toggle (or node name) changes.
+[metadata.settings.kubelet-dra-drivers.nvidia.enabled]
+affected-services = ["nvidia-dra-driver-gpu"]

variants/aws-k8s-1.34-nvidia/Cargo.toml

@@ -42,6 +42,7 @@ included-packages = [
     # nvidia
     "nvidia-container-toolkit-k8s",
     "nvidia-k8s-device-plugin",
+    "nvidia-dra-driver-gpu",
     "kmod-6.12-nvidia-r580-tesla",
 ]
 kernel-parameters = [

variants/aws-k8s-1.35-nvidia/Cargo.toml

@@ -42,6 +42,7 @@ included-packages = [
     # nvidia
     "nvidia-container-toolkit-k8s",
     "nvidia-k8s-device-plugin",
+    "nvidia-dra-driver-gpu",
     "kmod-6.12-nvidia-r580-tesla",
 ]
 kernel-parameters = [

variants/aws-k8s-1.36-nvidia/Cargo.toml

@@ -42,6 +42,7 @@ included-packages = [
     # nvidia
     "nvidia-container-toolkit-k8s",
     "nvidia-k8s-device-plugin",
+    "nvidia-dra-driver-gpu",
     "kmod-6.18-nvidia-r580-tesla",
 ]
 kernel-parameters = [