From 56adcd06f37a75696556b7a5209a4ca30cda6700 Mon Sep 17 00:00:00 2001
From: Patrick Ohly <patrick.ohly@intel.com>
Date: Thu, 20 Mar 2025 18:15:58 +0100
Subject: [PATCH] DRA device eviction: fix eviction triggered by pod scheduling

Normally the scheduler shouldn't schedule when there is a taint, but perhaps it
didn't know yet.

The TestEviction/update test covered this, but only failed under the right
timing conditions. The new event handler test case covers it reliably.
---
 .../device_taint_eviction.go                  |  3 ++-
 .../device_taint_eviction_test.go             | 21 +++++++++++++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/pkg/controller/devicetainteviction/device_taint_eviction.go b/pkg/controller/devicetainteviction/device_taint_eviction.go
index 31da9712e8e..3e2abb194d7 100644
--- a/pkg/controller/devicetainteviction/device_taint_eviction.go
+++ b/pkg/controller/devicetainteviction/device_taint_eviction.go
@@ -794,12 +794,13 @@ func (tc *Controller) handlePodChange(oldPod, newPod *v1.Pod) {
 
 	// Pods get updated quite frequently. There's no need
 	// to check them again unless something changed regarding
-	// their claims.
+	// their claims or they got scheduled.
 	//
 	// In particular this prevents adding the pod again
 	// directly after the eviction condition got added
 	// to it.
 	if oldPod != nil &&
+		oldPod.Spec.NodeName == newPod.Spec.NodeName &&
 		apiequality.Semantic.DeepEqual(oldPod.Status.ResourceClaimStatuses, newPod.Status.ResourceClaimStatuses) {
 		return
 	}
diff --git a/pkg/controller/devicetainteviction/device_taint_eviction_test.go b/pkg/controller/devicetainteviction/device_taint_eviction_test.go
index 60f7f2b154d..5ef0d8c25e6 100644
--- a/pkg/controller/devicetainteviction/device_taint_eviction_test.go
+++ b/pkg/controller/devicetainteviction/device_taint_eviction_test.go
@@ -318,6 +318,10 @@ var (
 			OwnerReference(podName, podUID+"-other", podKind).
 			UID("other").
 			Obj()
+	unscheduledPodWithClaimName = st.MakePod().Name(podName).Namespace(namespace).
+					UID(podUID).
+					PodResourceClaims(v1.PodResourceClaim{Name: resourceName, ResourceClaimName: &claimName}).
+					Obj()
 	podWithClaimName = st.MakePod().Name(podName).Namespace(namespace).
 				UID(podUID).
 				PodResourceClaims(v1.PodResourceClaim{Name: resourceName, ResourceClaimName: &claimName}).
@@ -494,6 +498,23 @@ func TestHandlers(t *testing.T) {
 			// At the moment, the code reliably cancels right away.
 			wantEvents: []*v1.Event{cancelPodEviction},
 		},
+		"evict-pod-after-scheduling": {
+			initialState: state{
+				pods:            []*v1.Pod{unscheduledPodWithClaimName},
+				slices:          []*resourceapi.ResourceSlice{sliceTainted, slice2},
+				allocatedClaims: []allocatedClaim{{ResourceClaim: inUseClaim, evictionTime: &taintTime}},
+			},
+			events: []any{
+				// Normally the scheduler shouldn't schedule when there is a taint,
+				// but perhaps it didn't know yet.
+				update(unscheduledPodWithClaimName, podWithClaimName),
+			},
+			finalState: state{
+				slices:          []*resourceapi.ResourceSlice{sliceTainted, slice2},
+				allocatedClaims: []allocatedClaim{{ResourceClaim: inUseClaim, evictionTime: &taintTime}},
+				evicting:        []evictAt{{newObject(podWithClaimName), taintTime.Time}},
+			},
+		},
 		"evict-pod-resourceclaim-unrelated-changes": {
 			initialState: state{
 				pods:            []*v1.Pod{podWithClaimName},