OCPBUGS-57520: e2e: llc: initial tests for cpu allocation

test/e2e/performanceprofile/functests/13_llc/llc.go

@@ -7,17 +7,23 @@ import (
 	"fmt"
 	"sort"
 	"strconv"
+	"strings"
 	"time"
 
 	. "github.com/onsi/ginkgo/v2"
 	. "github.com/onsi/gomega"
 
 	appsv1 "k8s.io/api/apps/v1"
 	corev1 "k8s.io/api/core/v1"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
 	"k8s.io/apimachinery/pkg/api/resource"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/fields"
 	"k8s.io/apimachinery/pkg/labels"
+
+	"github.com/jaypipes/ghw/pkg/cpu"
+	"github.com/jaypipes/ghw/pkg/topology"
+
 	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/utils/cpuset"
 	"k8s.io/utils/ptr"
@@ -27,6 +33,7 @@ import (
 	machineconfigv1 "github.com/openshift/api/machineconfiguration/v1"
 
 	performancev2 "github.com/openshift/cluster-node-tuning-operator/pkg/apis/performanceprofile/v2"
+	"github.com/openshift/cluster-node-tuning-operator/pkg/performanceprofile/controller/performanceprofile/components"
 	profilecomponent "github.com/openshift/cluster-node-tuning-operator/pkg/performanceprofile/controller/performanceprofile/components/profile"
 	testutils "github.com/openshift/cluster-node-tuning-operator/test/e2e/performanceprofile/functests/utils"
 	"github.com/openshift/cluster-node-tuning-operator/test/e2e/performanceprofile/functests/utils/cgroup"
@@ -41,6 +48,7 @@ import (
 	"github.com/openshift/cluster-node-tuning-operator/test/e2e/performanceprofile/functests/utils/nodes"
 	"github.com/openshift/cluster-node-tuning-operator/test/e2e/performanceprofile/functests/utils/pods"
 	"github.com/openshift/cluster-node-tuning-operator/test/e2e/performanceprofile/functests/utils/poolname"
+
 	"github.com/openshift/cluster-node-tuning-operator/test/e2e/performanceprofile/functests/utils/profiles"
 	"github.com/openshift/cluster-node-tuning-operator/test/e2e/performanceprofile/functests/utils/profilesupdate"
 )
@@ -54,9 +62,35 @@ const (
 	deploymentDeletionTime       = 1 * time.Minute
 )
 
+const (
+	// random number corresponding to the minimum we need. No known supported hardware has groups so little, they are all way bigger
+	expectedMinL3GroupSize = 8
+)
+
+type Machine struct {
+	CPU      *cpu.Info      `json:"cpu"`
+	Topology *topology.Info `json:"topology"`
+}
+
+type CacheInfo struct {
+	NodeID int
+	Level  int
+	CPUs   cpuset.CPUSet
+}
+
+func (ci CacheInfo) String() string {
+	return fmt.Sprintf("NUMANode=%d cacheLevel=%d cpus=<%s>", ci.NodeID, ci.Level, ci.CPUs.String())
+}
+
+type MachineData struct {
+	Info   Machine
+	Caches []CacheInfo
+}
+
 var _ = Describe("[rfe_id:77446] LLC-aware cpu pinning", Label(string(label.OpenShift), string(label.UnCoreCache)), Ordered, func() {
 	var (
 		workerRTNodes               []corev1.Node
+		machineDatas                map[string]MachineData // nodeName -> MachineData
 		initialProfile, perfProfile *performancev2.PerformanceProfile
 		performanceMCP              string
 		err                         error
@@ -69,26 +103,43 @@ var _ = Describe("[rfe_id:77446] LLC-aware cpu pinning", Label(string(label.Open
 	)
 
 	BeforeAll(func() {
+		var hasMachineData bool
 		profileAnnotations = make(map[string]string)
 		ctx := context.Background()
 
 		workerRTNodes, err = nodes.GetByLabels(testutils.NodeSelectorLabels)
 		Expect(err).ToNot(HaveOccurred())
 
 		workerRTNodes, err = nodes.MatchingOptionalSelector(workerRTNodes)
-		Expect(err).ToNot(HaveOccurred(), fmt.Sprintf("error looking for the optional selector: %v", err))
+		Expect(err).ToNot(HaveOccurred(), "error looking for the optional selector: %v", err)
+
+		if len(workerRTNodes) < 1 {
+			Skip("need at least a worker node")
+		}
 
 		perfProfile, err = profiles.GetByNodeLabels(testutils.NodeSelectorLabels)
 		Expect(err).ToNot(HaveOccurred())
 		initialProfile = perfProfile.DeepCopy()
 
+		By(fmt.Sprintf("collecting machine infos for %d nodes", len(workerRTNodes)))
+		machineDatas, hasMachineData, err = collectMachineDatas(ctx, workerRTNodes)
+		if !hasMachineData {
+			Skip("need machinedata available - please check the image for the presence of the machineinfo tool")
+		}
+		Expect(err).ToNot(HaveOccurred())
+
+		for node, data := range machineDatas {
+			testlog.Infof("node=%q data=%v", node, data.Caches)
+		}
+
 		getter, err = cgroup.BuildGetter(ctx, testclient.DataPlaneClient, testclient.K8sClient)
 		Expect(err).ToNot(HaveOccurred())
 		cgroupV2, err = cgroup.IsVersion2(ctx, testclient.DataPlaneClient)
 		Expect(err).ToNot(HaveOccurred())
 		if !cgroupV2 {
 			Skip("prefer-align-cpus-by-uncorecache cpumanager policy options is supported in cgroupv2 configuration only")
 		}
+
 		performanceMCP, err = mcps.GetByProfile(perfProfile)
 		Expect(err).ToNot(HaveOccurred())
 
@@ -116,16 +167,17 @@ var _ = Describe("[rfe_id:77446] LLC-aware cpu pinning", Label(string(label.Open
 		// Apply Annotation to enable align-cpu-by-uncorecache cpumanager policy option
 		if perfProfile.Annotations == nil || perfProfile.Annotations["kubeletconfig.experimental"] != llcPolicy {
 			testlog.Info("Enable align-cpus-by-uncorecache cpumanager policy")
-			perfProfile.Annotations = profileAnnotations
+			prof := perfProfile.DeepCopy()
+			prof.Annotations = profileAnnotations
 
 			By("updating performance profile")
-			profiles.UpdateWithRetry(perfProfile)
+			profiles.UpdateWithRetry(prof)
 
 			By(fmt.Sprintf("Applying changes in performance profile and waiting until %s will start updating", poolName))
-			profilesupdate.WaitForTuningUpdating(ctx, perfProfile)
+			profilesupdate.WaitForTuningUpdating(ctx, prof)
 
 			By(fmt.Sprintf("Waiting when %s finishes updates", poolName))
-			profilesupdate.WaitForTuningUpdated(ctx, perfProfile)
+			profilesupdate.WaitForTuningUpdated(ctx, prof)
 		}
 	})
 
@@ -215,6 +267,7 @@ var _ = Describe("[rfe_id:77446] LLC-aware cpu pinning", Label(string(label.Open
 			})
 		})
 	})
+
 	Context("Functional Tests with SMT Enabled", func() {
 		var (
 			L3CacheGroupSize int
@@ -854,6 +907,81 @@ var _ = Describe("[rfe_id:77446] LLC-aware cpu pinning", Label(string(label.Open
 			Entry("[test_id:81673] requesting less than L3Cache group size cpus", "test-deployment2", podCpuResourceLessthanL3CacheSize),
 		)
 	})
+
+	Context("Runtime Tests", func() {
+		var (
+			targetNodeName    string      // pick a random node to simplify our testing - e.g. to know ahead of time expected L3 group size
+			targetNodeInfo    MachineData // shortcut. Note: **SHALLOW COPY**
+			targetL3GroupSize int
+
+			testPod *corev1.Pod
+		)
+
+		BeforeEach(func() {
+			targetNodeName = workerRTNodes[0].Name // pick random node
+			var ok bool
+			targetNodeInfo, ok = machineDatas[targetNodeName]
+			Expect(ok).To(BeTrue(), "unknown machine data for node %q", targetNodeName)
+
+			targetL3GroupSize = expectedL3GroupSize(targetNodeInfo)
+			// random number corresponding to the minimum we need. No known supported hardware has groups so little, they are all way bigger
+			Expect(targetL3GroupSize).Should(BeNumerically(">", expectedMinL3GroupSize), "L3 Group size too small: %d", targetL3GroupSize)
+		})
+
+		// TODO move to DeferCleanup?
+		AfterEach(func() {
+			if testPod == nil {
+				return
+			}
+			ctx := context.Background()
+			testlog.Infof("deleting pod %q", testPod.Name)
+			deleteTestPod(ctx, testPod)
+		})
+
+		It("[test_id:77727] should align containers which request less than a L3 group size exclusive CPUs", func(ctx context.Context) {
+			askingCPUs := expectedMinL3GroupSize
+
+			By(fmt.Sprintf("Creating a test pod asking for %d exclusive CPUs", askingCPUs))
+			testPod = makeLLCPod(targetNodeName, askingCPUs)
+			Expect(testclient.Client.Create(ctx, testPod)).To(Succeed())
+
+			By("Waiting for the guaranteed pod to be ready")
+			testPod, err = pods.WaitForCondition(ctx, client.ObjectKeyFromObject(testPod), corev1.PodReady, corev1.ConditionTrue, 5*time.Minute)
+			Expect(err).ToNot(HaveOccurred(), "Guaranteed pod did not become ready in time")
+
+			logs, err := pods.GetLogs(testclient.K8sClient, testPod)
+			Expect(err).ToNot(HaveOccurred(), "Cannot get logs from test pod")
+
+			allocatedCPUs, err := cpuset.Parse(logs)
+			Expect(err).ToNot(HaveOccurred(), "Cannot get cpuset for pod %s/%s from logs %q", testPod.Namespace, testPod.Name, logs)
+			Expect(allocatedCPUs.Size()).To(Equal(askingCPUs), "asked %d exclusive CPUs got %v", askingCPUs, allocatedCPUs)
+
+			ok, _ := isCPUSetLLCAligned(targetNodeInfo.Caches, allocatedCPUs)
+			Expect(ok).To(BeTrue(), "pod has not L3-aligned CPUs") // TODO log what?
+		})
+
+		It("[test_id:81669] cannot align containers which request more than a L3 group size exclusive CPUs", func(ctx context.Context) {
+			askingCPUs := targetL3GroupSize + 2 // TODO: to be really safe we should add SMT level cpus
+
+			By(fmt.Sprintf("Creating a test pod asking for %d exclusive CPUs", askingCPUs))
+			testPod = makeLLCPod(targetNodeName, askingCPUs)
+			Expect(testclient.Client.Create(ctx, testPod)).To(Succeed())
+
+			By("Waiting for the guaranteed pod to be ready")
+			testPod, err = pods.WaitForCondition(ctx, client.ObjectKeyFromObject(testPod), corev1.PodReady, corev1.ConditionTrue, 5*time.Minute)
+			Expect(err).ToNot(HaveOccurred(), "Guaranteed pod did not become ready in time")
+
+			logs, err := pods.GetLogs(testclient.K8sClient, testPod)
+			Expect(err).ToNot(HaveOccurred(), "Cannot get logs from test pod")
+
+			allocatedCPUs, err := cpuset.Parse(logs)
+			Expect(err).ToNot(HaveOccurred(), "Cannot get cpuset for pod %s/%s from logs %q", testPod.Namespace, testPod.Name, logs)
+			Expect(allocatedCPUs.Size()).To(Equal(askingCPUs), "asked %d exclusive CPUs got %v", askingCPUs, allocatedCPUs)
+
+			ok, _ := isCPUSetLLCAligned(targetNodeInfo.Caches, allocatedCPUs)
+			Expect(ok).To(BeFalse(), "pod exceeds L3 group capacity so it cannot have L3-aligned CPUs") // TODO log what?
+		})
+	})
 })
 
 // create Machine config to create text file required to enable prefer-align-cpus-by-uncorecache policy option
@@ -995,3 +1123,137 @@ func transformToNoSMT(input map[int]map[int][]int) map[int][]int {
 	}
 	return result
 }
+
+func MachineFromJSON(data string) (Machine, error) {
+	ma := Machine{}
+	rd := strings.NewReader(data)
+	err := json.NewDecoder(rd).Decode(&ma)
+	return ma, err
+}
+
+func isCPUSetLLCAligned(infos []CacheInfo, cpus cpuset.CPUSet) (bool, *CacheInfo) {
+	for idx := range infos {
+		info := &infos[idx]
+		if cpus.IsSubsetOf(info.CPUs) {
+			return true, info
+		}
+	}
+	return false, nil
+}
+
+func computeLLCLayout(mi Machine) []CacheInfo {
+	ret := []CacheInfo{}
+	for _, node := range mi.Topology.Nodes {
+		for _, cache := range node.Caches {
+			if cache.Level < 3 { // TODO
+				continue
+			}
+			ret = append(ret, CacheInfo{
+				NodeID: node.ID,
+				Level:  int(cache.Level),
+				CPUs:   cpusetFromLogicalProcessors(cache.LogicalProcessors...),
+			})
+		}
+	}
+	return ret
+}
+
+func cpusetFromLogicalProcessors(procs ...uint32) cpuset.CPUSet {
+	cpuList := make([]int, 0, len(procs))
+	for _, proc := range procs {
+		cpuList = append(cpuList, int(proc))
+	}
+	return cpuset.New(cpuList...)
+}
+
+func expectedL3GroupSize(md MachineData) int {
+	// TODO: we assume all L3 Groups are equal in size.
+	for idx := range md.Caches {
+		cache := &md.Caches[idx]
+		if cache.Level != 3 {
+			continue
+		}
+		return cache.CPUs.Size()
+	}
+	return 0
+}
+
+func collectMachineDatas(ctx context.Context, nodeList []corev1.Node) (map[string]MachineData, bool, error) {
+	cmd := []string{"/usr/bin/machineinfo"}
+	infos := make(map[string]MachineData)
+	for idx := range nodeList {
+		node := &nodeList[idx]
+		out, err := nodes.ExecCommand(ctx, node, cmd)
+		if err != nil {
+			return infos, false, err
+		}
+
+		info, err := MachineFromJSON(string(out))
+		if err != nil {
+			return infos, true, err
+		}
+
+		infos[node.Name] = MachineData{
+			Info:   info,
+			Caches: computeLLCLayout(info), // precompute
+		}
+	}
+	return infos, true, nil
+}
+
+func makeLLCPod(nodeName string, guaranteedCPUs int) *corev1.Pod {
+	testPod := &corev1.Pod{
+		ObjectMeta: metav1.ObjectMeta{
+			GenerateName: "test-",
+			Labels: map[string]string{
+				"test": "",
+			},
+			Namespace: testutils.NamespaceTesting,
+		},
+		Spec: corev1.PodSpec{
+			Containers: []corev1.Container{
+				{
+					Name:  "test",
+					Image: images.Test(),
+					Command: []string{
+						"/bin/sh", "-c", "cat /sys/fs/cgroup/cpuset.cpus.effective && sleep 10h",
+					},
+				},
+			},
+			NodeName: nodeName,
+			NodeSelector: map[string]string{
+				testutils.LabelHostname: nodeName,
+			},
+		},
+	}
+	if guaranteedCPUs > 0 {
+		testPod.Spec.Containers[0].Resources = corev1.ResourceRequirements{
+			Limits: corev1.ResourceList{
+				corev1.ResourceCPU:    *resource.NewQuantity(int64(guaranteedCPUs), resource.DecimalSI),
+				corev1.ResourceMemory: resource.MustParse("256Mi"),
+			},
+		}
+	}
+	profile, _ := profiles.GetByNodeLabels(testutils.NodeSelectorLabels)
+	runtimeClass := components.GetComponentName(profile.Name, components.ComponentNamePrefix)
+	testPod.Spec.RuntimeClassName = &runtimeClass
+	return testPod
+}
+
+func deleteTestPod(ctx context.Context, testpod *corev1.Pod) bool {
+	GinkgoHelper()
+
+	// it possible that the pod already was deleted as part of the test, in this case we want to skip teardown
+	err := testclient.DataPlaneClient.Get(ctx, client.ObjectKeyFromObject(testpod), testpod)
+	if apierrors.IsNotFound(err) {
+		return false
+	}
+
+	err = testclient.DataPlaneClient.Delete(ctx, testpod)
+	Expect(err).ToNot(HaveOccurred())
+
+	err = pods.WaitForDeletion(ctx, testpod, pods.DefaultDeletionTimeout*time.Second)
+	Expect(err).ToNot(HaveOccurred())
+
+	return true
+}