30040328974

Committed 23 Jul 2026 08:00PM UTC coverage: 33.236% (+1.3%) from 31.955%

Build # 30040328974

Build Type

Pull #2571

github

Committed by

karthikvetrivel

Commit Message

Add golden render tests for GPUCluster operand manifests

Signed-off-by: Karthik Vetrivel <kvetrivel@nvidia.com>

Pull Request Pull Request #2571: Add GPUCluster CRD and controller for DRA-based stack

Coverage Stats

596 of 1510 new or added lines in 32 files covered. (39.47%)

18 existing lines in 4 files now uncovered.

4785 of 14397 relevant lines covered (33.24%)

0.38 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

52.66

/controllers/gpucluster_controller.go

/**
# Copyright (c) NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
**/

package controllers

import (
        "context"
        "fmt"
        "strings"
        "time"

        appsv1 "k8s.io/api/apps/v1"
        corev1 "k8s.io/api/core/v1"
        apierrors "k8s.io/apimachinery/pkg/api/errors"
        metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
        "k8s.io/apimachinery/pkg/runtime"
        "k8s.io/apimachinery/pkg/types"
        "k8s.io/client-go/tools/events"
        "k8s.io/client-go/util/workqueue"
        ctrl "sigs.k8s.io/controller-runtime"
        "sigs.k8s.io/controller-runtime/pkg/client"
        "sigs.k8s.io/controller-runtime/pkg/controller"
        "sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
        "sigs.k8s.io/controller-runtime/pkg/handler"
        "sigs.k8s.io/controller-runtime/pkg/log"
        "sigs.k8s.io/controller-runtime/pkg/predicate"
        "sigs.k8s.io/controller-runtime/pkg/reconcile"
        "sigs.k8s.io/controller-runtime/pkg/source"

        nvidiav1alpha1 "github.com/NVIDIA/gpu-operator/api/nvidia/v1alpha1"
        "github.com/NVIDIA/gpu-operator/controllers/clusterinfo"
        "github.com/NVIDIA/gpu-operator/internal/conditions"
        "github.com/NVIDIA/gpu-operator/internal/consts"
        "github.com/NVIDIA/gpu-operator/internal/state"
)

// gpuClusterFinalizer holds the GPUCluster until reconcileDelete has ordered teardown.
const gpuClusterFinalizer = "gpucluster.nvidia.com/dra-resourceclaim"

// draAdminNamespaceLabelKey is the label the kube-scheduler requires on a namespace
// before it allows adminAccess: true in ResourceClaim/ResourceClaimTemplate objects.
const draAdminNamespaceLabelKey = "resource.kubernetes.io/admin-access"

// GPUClusterReconciler reconciles a GPUCluster object
type GPUClusterReconciler struct {
        client.Client
        Scheme      *runtime.Scheme
        ClusterInfo clusterinfo.Interface
        Namespace   string

        stateManager     state.Manager
        conditionUpdater conditions.Updater
        recorder         events.EventRecorder

        // singleton is the GPUCluster that owns reconciliation; the first instance to
        // reconcile claims it (first-wins), mirroring ClusterPolicy.
        singleton *nvidiav1alpha1.GPUCluster
}

//+kubebuilder:rbac:groups=nvidia.com,resources=gpuclusters,verbs=get;list;watch;create;update;patch;delete
//+kubebuilder:rbac:groups=nvidia.com,resources=gpuclusters/status,verbs=get;update;patch
//+kubebuilder:rbac:groups=nvidia.com,resources=gpuclusters/finalizers,verbs=update
//+kubebuilder:rbac:groups=nvidia.com,resources=clusterpolicies,verbs=get;list;watch
//+kubebuilder:rbac:groups="",resources=namespaces,verbs=get;update;patch
//+kubebuilder:rbac:groups=events.k8s.io,resources=events,verbs=create;patch
//+kubebuilder:rbac:groups=resource.k8s.io,resources=resourceclaimtemplates,verbs=get;list;watch;create;update;delete

func (r *GPUClusterReconciler) Reconcile(ctx context.Context, req ctrl.Request) (ctrl.Result, error) {
        logger := log.FromContext(ctx)
        logger.V(consts.LogLevelInfo).Info("Reconciling GPUCluster")

        instance := &nvidiav1alpha1.GPUCluster{}
        if err := r.Get(ctx, req.NamespacedName, instance); err != nil {
                if apierrors.IsNotFound(err) {
                        // Deleted; owned objects are garbage-collected, so there is nothing to clean up.
                        return ctrl.Result{}, nil
                }
                // instance was not populated by the failed Get, so there is no object to
                // update status on; just surface the error for requeue.
                logger.Error(err, "error getting GPUCluster object")
                return ctrl.Result{}, fmt.Errorf("error getting GPUCluster object: %w", err)
        }

        if !instance.DeletionTimestamp.IsZero() {
                return r.reconcileDelete(ctx, instance)
        }
        if !controllerutil.ContainsFinalizer(instance, gpuClusterFinalizer) {
                controllerutil.AddFinalizer(instance, gpuClusterFinalizer)
                if err := r.Update(ctx, instance); err != nil {
                        return ctrl.Result{}, fmt.Errorf("error adding finalizer: %w", err)
                }
        }

        // GPUCluster (DRA stack) may coexist with a ClusterPolicy (device-plugin
        // stack): every operand DaemonSet of both stacks gates on the per-node
        // nvidia.com/gpu-operator.resource-allocation.mode label, so each node is served by exactly one stack.

        // Singleton, first-wins (mirroring ClusterPolicy): the first instance to reconcile
        // claims ownership; any other instance is marked Ignored and skipped. The owner is
        // held in memory, so the choice resets on operator restart.
        if r.singleton != nil && r.singleton.Name != instance.Name {
                logger.V(consts.LogLevelWarning).Info("Multiple GPUCluster instances found, ignoring this one",
                        "name", instance.Name, "owner", r.singleton.Name)
                if err := r.updateCRStatus(ctx, instance, nvidiav1alpha1.Ignored); err != nil {
                        return ctrl.Result{}, err
                }
                return ctrl.Result{}, nil
        }
        r.singleton = instance

        // The operand states render ResourceClaimTemplates with adminAccess: true, which the
        // kube-scheduler only admits from a labeled namespace; label it before syncing states.
        if err := r.ensureAdminAccessLabel(ctx); err != nil {
                return ctrl.Result{}, fmt.Errorf("failed to label namespace for admin access: %w", err)
        }

        infoCatalog := state.NewInfoCatalog()
        infoCatalog.Add(state.InfoTypeClusterInfo, r.ClusterInfo)

        managerStatus := r.stateManager.SyncState(ctx, instance, infoCatalog)

        if err := r.updateCRStatus(ctx, instance, nvidiav1alpha1.State(managerStatus.Status)); err != nil {
                return ctrl.Result{}, err
        }

        if managerStatus.Status != state.SyncStateReady {
                logger.Info("GPUCluster instance is not ready")
                for _, result := range managerStatus.StatesStatus {
                        if result.Status != state.SyncStateReady && result.ErrInfo != nil {
                                if condErr := r.conditionUpdater.SetConditionsError(ctx, instance, conditions.ReconcileFailed, fmt.Sprintf("Error syncing state %s: %v", result.StateName, result.ErrInfo)); condErr != nil {
                                        logger.Error(condErr, "failed to set condition")
                                }
                                return ctrl.Result{RequeueAfter: time.Second * 5}, nil
                        }
                }
                // no state reported an error, so we are waiting on operand pods
                if condErr := r.conditionUpdater.SetConditionsError(ctx, instance, conditions.OperandNotReady, "Waiting for operand pods to be ready"); condErr != nil {
                        logger.Error(condErr, "failed to set condition")
                }
                return ctrl.Result{RequeueAfter: time.Second * 5}, nil
        }

        if condErr := r.conditionUpdater.SetConditionsReady(ctx, instance, conditions.Reconciled, "All resources have been successfully reconciled"); condErr != nil {
                logger.Error(condErr, "failed to set condition")
                return ctrl.Result{}, condErr
        }
        // Resync periodically so out-of-band changes (a deleted DeviceClass/VAP, or a
        // newly-created ClusterPolicy) are detected and reconciled even while ready;
        // only DaemonSets are watched, and the ready path is otherwise event-driven.
        return ctrl.Result{RequeueAfter: time.Minute}, nil
}

// ensureAdminAccessLabel patches the operator namespace with the label required by the
// kube-scheduler to allow adminAccess: true in ResourceClaim/ResourceClaimTemplate
// objects. The label is deliberately never removed: it is namespace-level configuration
// that other adminAccess consumers in the namespace may rely on.
func (r *GPUClusterReconciler) ensureAdminAccessLabel(ctx context.Context) error {
        ns := &corev1.Namespace{}
        if err := r.Get(ctx, client.ObjectKey{Name: r.Namespace}, ns); err != nil {
                return fmt.Errorf("could not get namespace %s: %w", r.Namespace, err)
        }
        if ns.Labels[draAdminNamespaceLabelKey] == "true" {
                return nil
        }
        patch := client.MergeFrom(ns.DeepCopy())
        if ns.Labels == nil {
                ns.Labels = make(map[string]string)
        }
        ns.Labels[draAdminNamespaceLabelKey] = "true"
        return r.Patch(ctx, ns, patch)
}

// reconcileDelete drains ResourceClaim-consuming DaemonSets before releasing the CR:
// garbage collection would otherwise delete the DRA kubelet plugin while their pods
// still need it to unprepare claims, leaving them stuck in Terminating. Foreground
// propagation keeps each DaemonSet present until its pods are gone, i.e. unprepared.
func (r *GPUClusterReconciler) reconcileDelete(ctx context.Context, instance *nvidiav1alpha1.GPUCluster) (ctrl.Result, error) {
        logger := log.FromContext(ctx)

        if !controllerutil.ContainsFinalizer(instance, gpuClusterFinalizer) {
                return ctrl.Result{}, nil
        }

        dsList := &appsv1.DaemonSetList{}
        if err := r.List(ctx, dsList, client.InNamespace(r.Namespace)); err != nil {
                return ctrl.Result{}, fmt.Errorf("error listing DaemonSets: %w", err)
        }
        var draining []string
        for i := range dsList.Items {
                ds := &dsList.Items[i]
                if !metav1.IsControlledBy(ds, instance) || len(ds.Spec.Template.Spec.ResourceClaims) == 0 {
                        continue
                }
                draining = append(draining, ds.Name)
                if ds.DeletionTimestamp.IsZero() {
                        logger.V(consts.LogLevelInfo).Info("Draining ResourceClaim-consuming DaemonSet before teardown", "DaemonSet", ds.Name)
                        if err := r.Delete(ctx, ds, client.PropagationPolicy(metav1.DeletePropagationForeground)); err != nil && !apierrors.IsNotFound(err) {
                                return ctrl.Result{}, fmt.Errorf("error deleting DaemonSet %s: %w", ds.Name, err)
                        }
                }
        }
        if len(draining) > 0 {
                r.recorder.Eventf(instance, nil, corev1.EventTypeNormal, "DrainingClaimConsumers", "Delete",
                        "Waiting for ResourceClaim-consuming DaemonSet(s) to terminate: %s", strings.Join(draining, ", "))
                return ctrl.Result{RequeueAfter: time.Second * 5}, nil
        }

        controllerutil.RemoveFinalizer(instance, gpuClusterFinalizer)
        if err := r.Update(ctx, instance); err != nil {
                return ctrl.Result{}, fmt.Errorf("error removing finalizer: %w", err)
        }
        return ctrl.Result{}, nil
}

// updateCRStatus persists the given state (and the operator namespace) to the GPUCluster's
// .status subresource. It refetches the CR first to avoid resourceVersion conflicts and skips
// the API write when the status is already current. The desired status is mirrored onto cr
// up front so it is set on every non-error path.
func (r *GPUClusterReconciler) updateCRStatus(ctx context.Context, cr *nvidiav1alpha1.GPUCluster, desired nvidiav1alpha1.State) error {
        reqLogger := log.FromContext(ctx)

        // Refetch to avoid a resourceVersion conflict.
        instance := &nvidiav1alpha1.GPUCluster{}
        if err := r.Get(ctx, types.NamespacedName{Name: cr.Name}, instance); err != nil {
                reqLogger.Error(err, "Failed to get GPUCluster instance for status update")
                return err
        }
        cr.Status.State = desired
        cr.Status.Namespace = r.Namespace

        if instance.Status.State == desired && instance.Status.Namespace == r.Namespace {
                return nil
        }
        instance.Status.State = desired
        instance.Status.Namespace = r.Namespace

        reqLogger.V(consts.LogLevelInfo).Info("Updating CR Status", "Status", instance.Status)
        if err := r.Status().Update(ctx, instance); err != nil {
                reqLogger.Error(err, "Failed to update CR status")
                return err
        }
        return nil
}

// enqueueAllGPUClusters enqueues every instance so each is reconciled when any
// instance or owned resource changes.
func (r *GPUClusterReconciler) enqueueAllGPUClusters(ctx context.Context, _ *nvidiav1alpha1.GPUCluster) []reconcile.Request {
        logger := log.FromContext(ctx)
        list := &nvidiav1alpha1.GPUClusterList{}

        if err := r.List(ctx, list); err != nil {
                logger.Error(err, "Unable to list GPUCluster resources")
                return []reconcile.Request{}
        }

        reconcileRequests := make([]reconcile.Request, 0, len(list.Items))
        for _, config := range list.Items {
                reconcileRequests = append(reconcileRequests,
                        reconcile.Request{
                                NamespacedName: types.NamespacedName{
                                        Name: config.GetName(),
                                },
                        })
        }

        return reconcileRequests
}

func (r *GPUClusterReconciler) SetupWithManager(ctx context.Context, mgr ctrl.Manager) error {
        // The state manager renders the DRA driver operand for the GPUCluster.
        stateManager, err := state.NewManager(
                nvidiav1alpha1.GPUClusterCRDName,
                r.Namespace,
                mgr.GetClient(),
                mgr.GetScheme())
        if err != nil {
                return fmt.Errorf("error creating state manager: %w", err)
        }
        r.stateManager = stateManager

        r.conditionUpdater = conditions.NewGPUClusterUpdater(mgr.GetClient())
        r.recorder = mgr.GetEventRecorder("nvidia-gpu-operator")

        c, err := controller.New("gpu-cluster-controller", mgr, controller.Options{
                Reconciler:              r,
                MaxConcurrentReconciles: 1,
                RateLimiter:             workqueue.NewTypedItemExponentialFailureRateLimiter[reconcile.Request](minDelayCR, maxDelayCR),
        })
        if err != nil {
                return err
        }

        err = c.Watch(source.Kind(
                mgr.GetCache(),
                &nvidiav1alpha1.GPUCluster{},
                handler.TypedEnqueueRequestsFromMapFunc(r.enqueueAllGPUClusters),
                predicate.TypedGenerationChangedPredicate[*nvidiav1alpha1.GPUCluster]{},
        ),
        )
        if err != nil {
                return err
        }

        // Watch the secondary resources each state manager owns.
        watchSources := stateManager.GetWatchSources(mgr)
        for _, watchSource := range watchSources {
                err = c.Watch(
                        watchSource,
                )
                if err != nil {
                        return fmt.Errorf("error setting up Watch for source type %v: %w", watchSource, err)
                }
        }

        return nil
}

1	/**
2	# Copyright (c) NVIDIA CORPORATION. All rights reserved.
3	#
4	# Licensed under the Apache License, Version 2.0 (the "License");
5	# you may not use this file except in compliance with the License.
6	# You may obtain a copy of the License at
7	#
8	# http://www.apache.org/licenses/LICENSE-2.0
9	#
10	# Unless required by applicable law or agreed to in writing, software
11	# distributed under the License is distributed on an "AS IS" BASIS,
12	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13	# See the License for the specific language governing permissions and
14	# limitations under the License.
15	**/
16
17	package controllers
18
19	import (
20	"context"
21	"fmt"
22	"strings"
23	"time"
24
25	appsv1 "k8s.io/api/apps/v1"
26	corev1 "k8s.io/api/core/v1"
27	apierrors "k8s.io/apimachinery/pkg/api/errors"
28	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
29	"k8s.io/apimachinery/pkg/runtime"
30	"k8s.io/apimachinery/pkg/types"
31	"k8s.io/client-go/tools/events"
32	"k8s.io/client-go/util/workqueue"
33	ctrl "sigs.k8s.io/controller-runtime"
34	"sigs.k8s.io/controller-runtime/pkg/client"
35	"sigs.k8s.io/controller-runtime/pkg/controller"
36	"sigs.k8s.io/controller-runtime/pkg/controller/controllerutil"
37	"sigs.k8s.io/controller-runtime/pkg/handler"
38	"sigs.k8s.io/controller-runtime/pkg/log"
39	"sigs.k8s.io/controller-runtime/pkg/predicate"
40	"sigs.k8s.io/controller-runtime/pkg/reconcile"
41	"sigs.k8s.io/controller-runtime/pkg/source"
42
43	nvidiav1alpha1 "github.com/NVIDIA/gpu-operator/api/nvidia/v1alpha1"
44	"github.com/NVIDIA/gpu-operator/controllers/clusterinfo"
45	"github.com/NVIDIA/gpu-operator/internal/conditions"
46	"github.com/NVIDIA/gpu-operator/internal/consts"
47	"github.com/NVIDIA/gpu-operator/internal/state"
48	)
49
50	// gpuClusterFinalizer holds the GPUCluster until reconcileDelete has ordered teardown.
51	const gpuClusterFinalizer = "gpucluster.nvidia.com/dra-resourceclaim"
52
53	// draAdminNamespaceLabelKey is the label the kube-scheduler requires on a namespace
54	// before it allows adminAccess: true in ResourceClaim/ResourceClaimTemplate objects.
55	const draAdminNamespaceLabelKey = "resource.kubernetes.io/admin-access"
56
57	// GPUClusterReconciler reconciles a GPUCluster object
58	type GPUClusterReconciler struct {
59	client.Client
60	Scheme *runtime.Scheme
61	ClusterInfo clusterinfo.Interface
62	Namespace string
63
64	stateManager state.Manager
65	conditionUpdater conditions.Updater
66	recorder events.EventRecorder
67
68	// singleton is the GPUCluster that owns reconciliation; the first instance to
69	// reconcile claims it (first-wins), mirroring ClusterPolicy.
70	singleton *nvidiav1alpha1.GPUCluster
71	}
72
73	//+kubebuilder:rbac:groups=nvidia.com,resources=gpuclusters,verbs=get;list;watch;create;update;patch;delete
74	//+kubebuilder:rbac:groups=nvidia.com,resources=gpuclusters/status,verbs=get;update;patch
75	//+kubebuilder:rbac:groups=nvidia.com,resources=gpuclusters/finalizers,verbs=update
76	//+kubebuilder:rbac:groups=nvidia.com,resources=clusterpolicies,verbs=get;list;watch
77	//+kubebuilder:rbac:groups="",resources=namespaces,verbs=get;update;patch
78	//+kubebuilder:rbac:groups=events.k8s.io,resources=events,verbs=create;patch
79	//+kubebuilder:rbac:groups=resource.k8s.io,resources=resourceclaimtemplates,verbs=get;list;watch;create;update;delete
80
81	func (r *GPUClusterReconciler) Reconcile(ctx context.Context, req ctrl.Request) (ctrl.Result, error) {	1✔
82	logger := log.FromContext(ctx)	1✔
83	logger.V(consts.LogLevelInfo).Info("Reconciling GPUCluster")	1✔
84		1✔
85	instance := &nvidiav1alpha1.GPUCluster{}	1✔
86	if err := r.Get(ctx, req.NamespacedName, instance); err != nil {	2✔
87	if apierrors.IsNotFound(err) {	2✔
88	// Deleted; owned objects are garbage-collected, so there is nothing to clean up.	1✔
89	return ctrl.Result{}, nil	1✔
90	}	1✔
91	// instance was not populated by the failed Get, so there is no object to
92	// update status on; just surface the error for requeue.
NEW 93	logger.Error(err, "error getting GPUCluster object")	×
NEW 94	return ctrl.Result{}, fmt.Errorf("error getting GPUCluster object: %w", err)	×
95	}
96
97	if !instance.DeletionTimestamp.IsZero() {	2✔
98	return r.reconcileDelete(ctx, instance)	1✔
99	}	1✔
100	if !controllerutil.ContainsFinalizer(instance, gpuClusterFinalizer) {	2✔
101	controllerutil.AddFinalizer(instance, gpuClusterFinalizer)	1✔
102	if err := r.Update(ctx, instance); err != nil {	1✔
NEW 103	return ctrl.Result{}, fmt.Errorf("error adding finalizer: %w", err)	×
NEW 104	}	×
105	}
106
107	// GPUCluster (DRA stack) may coexist with a ClusterPolicy (device-plugin
108	// stack): every operand DaemonSet of both stacks gates on the per-node
109	// nvidia.com/gpu-operator.resource-allocation.mode label, so each node is served by exactly one stack.
110
111	// Singleton, first-wins (mirroring ClusterPolicy): the first instance to reconcile
112	// claims ownership; any other instance is marked Ignored and skipped. The owner is
113	// held in memory, so the choice resets on operator restart.
114	if r.singleton != nil && r.singleton.Name != instance.Name {	2✔
115	logger.V(consts.LogLevelWarning).Info("Multiple GPUCluster instances found, ignoring this one",	1✔
116	"name", instance.Name, "owner", r.singleton.Name)	1✔
117	if err := r.updateCRStatus(ctx, instance, nvidiav1alpha1.Ignored); err != nil {	1✔
NEW 118	return ctrl.Result{}, err	×
NEW 119	}	×
120	return ctrl.Result{}, nil	1✔
121	}
122	r.singleton = instance	1✔
123		1✔
124	// The operand states render ResourceClaimTemplates with adminAccess: true, which the	1✔
125	// kube-scheduler only admits from a labeled namespace; label it before syncing states.	1✔
126	if err := r.ensureAdminAccessLabel(ctx); err != nil {	1✔
NEW 127	return ctrl.Result{}, fmt.Errorf("failed to label namespace for admin access: %w", err)	×
NEW 128	}	×
129
130	infoCatalog := state.NewInfoCatalog()	1✔
131	infoCatalog.Add(state.InfoTypeClusterInfo, r.ClusterInfo)	1✔
132		1✔
133	managerStatus := r.stateManager.SyncState(ctx, instance, infoCatalog)	1✔
134		1✔
135	if err := r.updateCRStatus(ctx, instance, nvidiav1alpha1.State(managerStatus.Status)); err != nil {	1✔
NEW 136	return ctrl.Result{}, err	×
NEW 137	}	×
138
139	if managerStatus.Status != state.SyncStateReady {	1✔
NEW 140	logger.Info("GPUCluster instance is not ready")	×
NEW 141	for _, result := range managerStatus.StatesStatus {	×
NEW 142	if result.Status != state.SyncStateReady && result.ErrInfo != nil {	×
NEW 143	if condErr := r.conditionUpdater.SetConditionsError(ctx, instance, conditions.ReconcileFailed, fmt.Sprintf("Error syncing state %s: %v", result.StateName, result.ErrInfo)); condErr != nil {	×
NEW 144	logger.Error(condErr, "failed to set condition")	×
NEW 145	}	×
NEW 146	return ctrl.Result{RequeueAfter: time.Second * 5}, nil	×
147	}
148	}
149	// no state reported an error, so we are waiting on operand pods
NEW 150	if condErr := r.conditionUpdater.SetConditionsError(ctx, instance, conditions.OperandNotReady, "Waiting for operand pods to be ready"); condErr != nil {	×
NEW 151	logger.Error(condErr, "failed to set condition")	×
NEW 152	}	×
NEW 153	return ctrl.Result{RequeueAfter: time.Second * 5}, nil	×
154	}
155
156	if condErr := r.conditionUpdater.SetConditionsReady(ctx, instance, conditions.Reconciled, "All resources have been successfully reconciled"); condErr != nil {	1✔
NEW 157	logger.Error(condErr, "failed to set condition")	×
NEW 158	return ctrl.Result{}, condErr	×
NEW 159	}	×
160	// Resync periodically so out-of-band changes (a deleted DeviceClass/VAP, or a
161	// newly-created ClusterPolicy) are detected and reconciled even while ready;
162	// only DaemonSets are watched, and the ready path is otherwise event-driven.
163	return ctrl.Result{RequeueAfter: time.Minute}, nil	1✔
164	}
165
166	// ensureAdminAccessLabel patches the operator namespace with the label required by the
167	// kube-scheduler to allow adminAccess: true in ResourceClaim/ResourceClaimTemplate
168	// objects. The label is deliberately never removed: it is namespace-level configuration
169	// that other adminAccess consumers in the namespace may rely on.
170	func (r *GPUClusterReconciler) ensureAdminAccessLabel(ctx context.Context) error {	1✔
171	ns := &corev1.Namespace{}	1✔
172	if err := r.Get(ctx, client.ObjectKey{Name: r.Namespace}, ns); err != nil {	1✔
NEW 173	return fmt.Errorf("could not get namespace %s: %w", r.Namespace, err)	×
NEW 174	}	×
175	if ns.Labels[draAdminNamespaceLabelKey] == "true" {	1✔
NEW 176	return nil	×
NEW 177	}	×
178	patch := client.MergeFrom(ns.DeepCopy())	1✔
179	if ns.Labels == nil {	2✔
180	ns.Labels = make(map[string]string)	1✔
181	}	1✔
182	ns.Labels[draAdminNamespaceLabelKey] = "true"	1✔
183	return r.Patch(ctx, ns, patch)	1✔
184	}
185
186	// reconcileDelete drains ResourceClaim-consuming DaemonSets before releasing the CR:
187	// garbage collection would otherwise delete the DRA kubelet plugin while their pods
188	// still need it to unprepare claims, leaving them stuck in Terminating. Foreground
189	// propagation keeps each DaemonSet present until its pods are gone, i.e. unprepared.
190	func (r GPUClusterReconciler) reconcileDelete(ctx context.Context, instance nvidiav1alpha1.GPUCluster) (ctrl.Result, error) {	1✔
191	logger := log.FromContext(ctx)	1✔
192		1✔
193	if !controllerutil.ContainsFinalizer(instance, gpuClusterFinalizer) {	1✔
NEW 194	return ctrl.Result{}, nil	×
NEW 195	}	×
196
197	dsList := &appsv1.DaemonSetList{}	1✔
198	if err := r.List(ctx, dsList, client.InNamespace(r.Namespace)); err != nil {	1✔
NEW 199	return ctrl.Result{}, fmt.Errorf("error listing DaemonSets: %w", err)	×
NEW 200	}	×
201	var draining []string	1✔
202	for i := range dsList.Items {	2✔
203	ds := &dsList.Items[i]	1✔
204	if !metav1.IsControlledBy(ds, instance) \|\| len(ds.Spec.Template.Spec.ResourceClaims) == 0 {	2✔
205	continue	1✔
206	}
207	draining = append(draining, ds.Name)	1✔
208	if ds.DeletionTimestamp.IsZero() {	2✔
209	logger.V(consts.LogLevelInfo).Info("Draining ResourceClaim-consuming DaemonSet before teardown", "DaemonSet", ds.Name)	1✔
210	if err := r.Delete(ctx, ds, client.PropagationPolicy(metav1.DeletePropagationForeground)); err != nil && !apierrors.IsNotFound(err) {	1✔
NEW 211	return ctrl.Result{}, fmt.Errorf("error deleting DaemonSet %s: %w", ds.Name, err)	×
NEW 212	}	×
213	}
214	}
215	if len(draining) > 0 {	2✔
216	r.recorder.Eventf(instance, nil, corev1.EventTypeNormal, "DrainingClaimConsumers", "Delete",	1✔
217	"Waiting for ResourceClaim-consuming DaemonSet(s) to terminate: %s", strings.Join(draining, ", "))	1✔
218	return ctrl.Result{RequeueAfter: time.Second * 5}, nil	1✔
219	}	1✔
220
221	controllerutil.RemoveFinalizer(instance, gpuClusterFinalizer)	1✔
222	if err := r.Update(ctx, instance); err != nil {	1✔
NEW 223	return ctrl.Result{}, fmt.Errorf("error removing finalizer: %w", err)	×
NEW 224	}	×
225	return ctrl.Result{}, nil	1✔
226	}
227
228	// updateCRStatus persists the given state (and the operator namespace) to the GPUCluster's
229	// .status subresource. It refetches the CR first to avoid resourceVersion conflicts and skips
230	// the API write when the status is already current. The desired status is mirrored onto cr
231	// up front so it is set on every non-error path.
232	func (r GPUClusterReconciler) updateCRStatus(ctx context.Context, cr nvidiav1alpha1.GPUCluster, desired nvidiav1alpha1.State) error {	1✔
233	reqLogger := log.FromContext(ctx)	1✔
234		1✔
235	// Refetch to avoid a resourceVersion conflict.	1✔
236	instance := &nvidiav1alpha1.GPUCluster{}	1✔
237	if err := r.Get(ctx, types.NamespacedName{Name: cr.Name}, instance); err != nil {	1✔
NEW 238	reqLogger.Error(err, "Failed to get GPUCluster instance for status update")	×
NEW 239	return err	×
NEW 240	}	×
241	cr.Status.State = desired	1✔
242	cr.Status.Namespace = r.Namespace	1✔
243		1✔
244	if instance.Status.State == desired && instance.Status.Namespace == r.Namespace {	1✔
NEW 245	return nil	×
NEW 246	}	×
247	instance.Status.State = desired	1✔
248	instance.Status.Namespace = r.Namespace	1✔
249		1✔
250	reqLogger.V(consts.LogLevelInfo).Info("Updating CR Status", "Status", instance.Status)	1✔
251	if err := r.Status().Update(ctx, instance); err != nil {	1✔
NEW 252	reqLogger.Error(err, "Failed to update CR status")	×
NEW 253	return err	×
NEW 254	}	×
255	return nil	1✔
256	}
257
258	// enqueueAllGPUClusters enqueues every instance so each is reconciled when any
259	// instance or owned resource changes.
260	func (r GPUClusterReconciler) enqueueAllGPUClusters(ctx context.Context, _ nvidiav1alpha1.GPUCluster) []reconcile.Request {	1✔
261	logger := log.FromContext(ctx)	1✔
262	list := &nvidiav1alpha1.GPUClusterList{}	1✔
263		1✔
264	if err := r.List(ctx, list); err != nil {	1✔
NEW 265	logger.Error(err, "Unable to list GPUCluster resources")	×
NEW 266	return []reconcile.Request{}	×
NEW 267	}	×
268
269	reconcileRequests := make([]reconcile.Request, 0, len(list.Items))	1✔
270	for _, config := range list.Items {	2✔
271	reconcileRequests = append(reconcileRequests,	1✔
272	reconcile.Request{	1✔
273	NamespacedName: types.NamespacedName{	1✔
274	Name: config.GetName(),	1✔
275	},	1✔
276	})	1✔
277	}	1✔
278
279	return reconcileRequests	1✔
280	}
281
NEW 282	func (r *GPUClusterReconciler) SetupWithManager(ctx context.Context, mgr ctrl.Manager) error {	×
NEW 283	// The state manager renders the DRA driver operand for the GPUCluster.	×
NEW 284	stateManager, err := state.NewManager(	×
NEW 285	nvidiav1alpha1.GPUClusterCRDName,	×
NEW 286	r.Namespace,	×
NEW 287	mgr.GetClient(),	×
NEW 288	mgr.GetScheme())	×
NEW 289	if err != nil {	×
NEW 290	return fmt.Errorf("error creating state manager: %w", err)	×
NEW 291	}	×
NEW 292	r.stateManager = stateManager	×
NEW 293		×
NEW 294	r.conditionUpdater = conditions.NewGPUClusterUpdater(mgr.GetClient())	×
NEW 295	r.recorder = mgr.GetEventRecorder("nvidia-gpu-operator")	×
NEW 296		×
NEW 297	c, err := controller.New("gpu-cluster-controller", mgr, controller.Options{	×
NEW 298	Reconciler: r,	×
NEW 299	MaxConcurrentReconciles: 1,	×
NEW 300	RateLimiter: workqueue.NewTypedItemExponentialFailureRateLimiter[reconcile.Request](minDelayCR, maxDelayCR),	×
NEW 301	})	×
NEW 302	if err != nil {	×
NEW 303	return err	×
NEW 304	}	×
305
NEW 306	err = c.Watch(source.Kind(	×
NEW 307	mgr.GetCache(),	×
NEW 308	&nvidiav1alpha1.GPUCluster{},	×
NEW 309	handler.TypedEnqueueRequestsFromMapFunc(r.enqueueAllGPUClusters),	×
NEW 310	predicate.TypedGenerationChangedPredicate[*nvidiav1alpha1.GPUCluster]{},	×
NEW 311	),	×
NEW 312	)	×
NEW 313	if err != nil {	×
NEW 314	return err	×
NEW 315	}	×
316
317	// Watch the secondary resources each state manager owns.
NEW 318	watchSources := stateManager.GetWatchSources(mgr)	×
NEW 319	for _, watchSource := range watchSources {	×
NEW 320	err = c.Watch(	×
NEW 321	watchSource,	×
NEW 322	)	×
NEW 323	if err != nil {	×
NEW 324	return fmt.Errorf("error setting up Watch for source type %v: %w", watchSource, err)	×
NEW 325	}	×
326	}
327
NEW 328	return nil	×
329	}

NVIDIA / gpu-operator / 30040328974

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous