e2e: topomgr: check pod resource alignment

This patch extends and completes the previously-added
empty topology manager test for single-NUMA node policy
by adding reporting in the test pod and checking
the resource alignment.

Signed-off-by: Francesco Romani <fromani@redhat.com>

test/e2e_node/BUILD

@@ -15,6 +15,7 @@ go_library(
         "framework.go",
         "image_list.go",
         "node_problem_detector_linux.go",
+        "numa_alignment.go",
         "resource_collector.go",
         "util.go",
         "util_sriov.go",
@@ -31,6 +32,7 @@ go_library(
         "//pkg/kubelet/apis/podresources/v1alpha1:go_default_library",
         "//pkg/kubelet/apis/stats/v1alpha1:go_default_library",
         "//pkg/kubelet/cm:go_default_library",
+        "//pkg/kubelet/cm/cpuset:go_default_library",
         "//pkg/kubelet/kubeletconfig/util/codec:go_default_library",
         "//pkg/kubelet/metrics:go_default_library",
         "//pkg/kubelet/remote:go_default_library",

test/e2e_node/numa_alignment.go

@@ -0,0 +1,199 @@
+/*
+Copyright 2020 The Kubernetes Authors.
+
+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 e2enode
+
+import (
+	"fmt"
+	"sort"
+	"strconv"
+	"strings"
+
+	v1 "k8s.io/api/core/v1"
+	"k8s.io/kubernetes/pkg/kubelet/cm/cpuset"
+
+	"k8s.io/kubernetes/test/e2e/framework"
+)
+
+type numaPodResources struct {
+	CPUToNUMANode     map[int]int
+	PCIDevsToNUMANode map[string]int
+}
+
+func (R *numaPodResources) CheckAlignment() bool {
+	nodeNum := -1 // not set
+	for _, cpuNode := range R.CPUToNUMANode {
+		if nodeNum == -1 {
+			nodeNum = cpuNode
+		} else if nodeNum != cpuNode {
+			return false
+		}
+	}
+	for _, devNode := range R.PCIDevsToNUMANode {
+		// TODO: explain -1
+		if devNode != -1 && nodeNum != devNode {
+			return false
+		}
+	}
+	return true
+}
+
+func (R *numaPodResources) String() string {
+	var b strings.Builder
+	// To store the keys in slice in sorted order
+	var cpuKeys []int
+	for ck := range R.CPUToNUMANode {
+		cpuKeys = append(cpuKeys, ck)
+	}
+	sort.Ints(cpuKeys)
+	for _, k := range cpuKeys {
+		nodeNum := R.CPUToNUMANode[k]
+		b.WriteString(fmt.Sprintf("CPU cpu#%03d=%02d\n", k, nodeNum))
+	}
+	var pciKeys []string
+	for pk := range R.PCIDevsToNUMANode {
+		pciKeys = append(pciKeys, pk)
+	}
+	sort.Strings(pciKeys)
+	for _, k := range pciKeys {
+		nodeNum := R.PCIDevsToNUMANode[k]
+		b.WriteString(fmt.Sprintf("PCI %s=%02d\n", k, nodeNum))
+	}
+	return b.String()
+}
+
+func getCPUToNUMANodeMapFromEnv(f *framework.Framework, pod *v1.Pod, environ map[string]string, numaNodes int) (map[int]int, error) {
+	var cpuIDs []int
+	cpuListAllowedEnvVar := "CPULIST_ALLOWED"
+
+	for name, value := range environ {
+		if name == cpuListAllowedEnvVar {
+			cpus, err := cpuset.Parse(value)
+			if err != nil {
+				return nil, err
+			}
+			cpuIDs = cpus.ToSlice()
+		}
+	}
+	if len(cpuIDs) == 0 {
+		return nil, fmt.Errorf("variable %q found in environ", cpuListAllowedEnvVar)
+	}
+
+	cpusPerNUMA := make(map[int][]int)
+	for numaNode := 0; numaNode < numaNodes; numaNode++ {
+		nodeCPUList := f.ExecCommandInContainer(pod.Name, pod.Spec.Containers[0].Name,
+			"/bin/cat", fmt.Sprintf("/sys/devices/system/node/node%d/cpulist", numaNode))
+
+		cpus, err := cpuset.Parse(nodeCPUList)
+		if err != nil {
+			return nil, err
+		}
+		cpusPerNUMA[numaNode] = cpus.ToSlice()
+	}
+
+	// CPU IDs -> NUMA Node ID
+	CPUToNUMANode := make(map[int]int)
+	for nodeNum, cpus := range cpusPerNUMA {
+		for _, cpu := range cpus {
+			CPUToNUMANode[cpu] = nodeNum
+		}
+	}
+
+	// filter out only the allowed CPUs
+	CPUMap := make(map[int]int)
+	for _, cpuID := range cpuIDs {
+		_, ok := CPUToNUMANode[cpuID]
+		if !ok {
+			return nil, fmt.Errorf("CPU %d not found on NUMA map: %v", cpuID, CPUToNUMANode)
+		}
+		CPUMap[cpuID] = CPUToNUMANode[cpuID]
+	}
+	return CPUMap, nil
+}
+
+func getPCIDeviceToNumaNodeMapFromEnv(f *framework.Framework, pod *v1.Pod, environ map[string]string) (map[string]int, error) {
+	pciDevPrefix := "PCIDEVICE_"
+	// at this point we don't care which plugin selected the device,
+	// we only need to know which devices were assigned to the POD.
+	// Hence, do prefix search for the variable and fetch the device(s).
+
+	NUMAPerDev := make(map[string]int)
+	for name, value := range environ {
+		if !strings.HasPrefix(name, pciDevPrefix) {
+			continue
+		}
+
+		// a single plugin can allocate more than a single device
+		pciDevs := strings.Split(value, ",")
+		for _, pciDev := range pciDevs {
+			pciDevNUMANode := f.ExecCommandInContainer(pod.Name, pod.Spec.Containers[0].Name,
+				"/bin/cat", fmt.Sprintf("/sys/bus/pci/devices/%s/numa_node", pciDev))
+
+			nodeNum, err := strconv.Atoi(pciDevNUMANode)
+			if err != nil {
+				return nil, err
+			}
+			NUMAPerDev[pciDev] = nodeNum
+		}
+	}
+	if len(NUMAPerDev) == 0 {
+		return nil, fmt.Errorf("no PCI devices found in environ")
+	}
+	return NUMAPerDev, nil
+}
+
+func makeEnvMap(logs string) (map[string]string, error) {
+	podEnv := strings.Split(logs, "\n")
+	envMap := make(map[string]string)
+	for _, envVar := range podEnv {
+		if len(envVar) == 0 {
+			continue
+		}
+		pair := strings.SplitN(envVar, "=", 2)
+		if len(pair) != 2 {
+			return nil, fmt.Errorf("unable to split %q", envVar)
+		}
+		envMap[pair[0]] = pair[1]
+	}
+	return envMap, nil
+}
+
+func checkNUMAAlignment(f *framework.Framework, pod *v1.Pod, logs string, numaNodes int) (numaPodResources, error) {
+	podEnv, err := makeEnvMap(logs)
+	if err != nil {
+		return numaPodResources{}, err
+	}
+
+	CPUToNUMANode, err := getCPUToNUMANodeMapFromEnv(f, pod, podEnv, numaNodes)
+	if err != nil {
+		return numaPodResources{}, err
+	}
+
+	PCIDevsToNUMANode, err := getPCIDeviceToNumaNodeMapFromEnv(f, pod, podEnv)
+	if err != nil {
+		return numaPodResources{}, err
+	}
+
+	numaRes := numaPodResources{
+		CPUToNUMANode:     CPUToNUMANode,
+		PCIDevsToNUMANode: PCIDevsToNUMANode,
+	}
+	aligned := numaRes.CheckAlignment()
+	if !aligned {
+		return numaRes, fmt.Errorf("NUMA resources not aligned")
+	}
+	return numaRes, nil
+}

test/e2e_node/topology_manager_test.go

@@ -43,6 +43,10 @@ import (
 	"github.com/onsi/gomega"
 )
 
+const (
+	numalignCmd = `export CPULIST_ALLOWED=$( awk -F":\t*" '/Cpus_allowed_list/ { print $2 }' /proc/self/status); env; sleep 1d`
+)
+
 // Helper for makeTopologyManagerPod().
 type tmCtnAttribute struct {
 	ctnName     string
@@ -222,6 +226,18 @@ func deletePodInNamespace(f *framework.Framework, namespace, name string) {
 	framework.ExpectNoError(err)
 }
 
+func validatePodAlignment(f *framework.Framework, pod *v1.Pod, numaNodes int) {
+	ginkgo.By("validating the Gu pod")
+	logs, err := e2epod.GetPodLogs(f.ClientSet, f.Namespace.Name, pod.Name, pod.Spec.Containers[0].Name)
+	framework.ExpectNoError(err, "expected log not found in container [%s] of pod [%s]",
+		pod.Spec.Containers[0].Name, pod.Name)
+
+	framework.Logf("got pod logs: %v", logs)
+	numaRes, err := checkNUMAAlignment(f, pod, logs, numaNodes)
+	framework.ExpectNoError(err, "NUMA Alignment check failed for [%s] of pod [%s]: %s",
+		pod.Spec.Containers[0].Name, pod.Name, numaRes.String())
+}
+
 func runTopologyManagerPolicySuiteTests(f *framework.Framework) {
 	var cpuCap, cpuAlloc int64
 	var cpuListString, expAllowedCPUsListRegex string
@@ -448,7 +464,29 @@ func runTopologyManagerPolicySuiteTests(f *framework.Framework) {
 	waitForContainerRemoval(pod2.Spec.Containers[0].Name, pod2.Name, pod2.Namespace)
 }
 
-func runTopologyManagerNodeAlignmentSuiteTests(f *framework.Framework) {
+func runTopologyManagerNodeAlignmentSinglePodTest(f *framework.Framework, numaNodes int) {
+	ginkgo.By("allocate aligned resources for a single pod")
+	ctnAttrs := []tmCtnAttribute{
+		{
+			ctnName:     "gu-container",
+			cpuRequest:  "1000m",
+			cpuLimit:    "1000m",
+			devResource: SRIOVResourceName,
+		},
+	}
+
+	pod := makeTopologyManagerTestPod("gu-pod", numalignCmd, ctnAttrs)
+	pod = f.PodClient().CreateSync(pod)
+
+	validatePodAlignment(f, pod, numaNodes)
+
+	framework.Logf("deleting the pod %s/%s and waiting for container %s removal",
+		pod.Namespace, pod.Name, pod.Spec.Containers[0].Name)
+	deletePods(f, []string{pod.Name})
+	waitForContainerRemoval(pod.Spec.Containers[0].Name, pod.Name, pod.Namespace)
+}
+
+func runTopologyManagerNodeAlignmentSuiteTests(f *framework.Framework, numaNodes int) {
 	var err error
 
 	configMap := readConfigMapV1OrDie(testfiles.ReadOrDie(SRIOVDevicePluginCMYAML))
@@ -480,27 +518,7 @@ func runTopologyManagerNodeAlignmentSuiteTests(f *framework.Framework) {
 	}, 5*time.Minute, framework.Poll).Should(gomega.BeTrue())
 	framework.Logf("Successfully created device plugin pod")
 
-	ginkgo.By("running a Gu pod")
+	runTopologyManagerNodeAlignmentSinglePodTest(f, numaNodes)
-	ctnAttrs := []tmCtnAttribute{
-		{
-			ctnName:     "gu-container",
-			cpuRequest:  "1000m",
-			cpuLimit:    "1000m",
-			devResource: SRIOVResourceName,
-		},
-	}
-
-	pod := makeTopologyManagerTestPod("gu-pod", "env && sleep 1d", ctnAttrs)
-	pod = f.PodClient().CreateSync(pod)
-
-	ginkgo.By("validating the Gu pod")
-	_, err = e2epod.GetPodLogs(f.ClientSet, f.Namespace.Name, pod.Name, pod.Spec.Containers[0].Name)
-	framework.ExpectNoError(err, "expected log not found in container [%s] of pod [%s]",
-		pod.Spec.Containers[0].Name, pod.Name)
-
-	ginkgo.By("by deleting the pods and waiting for container removal")
-	deletePods(f, []string{pod.Name})
-	waitForContainerRemoval(pod.Spec.Containers[0].Name, pod.Name, pod.Namespace)
 
 	framework.Logf("deleting the SRIOV device plugin pod %s/%s and waiting for container %s removal",
 		dpPod.Namespace, dpPod.Name, dpPod.Spec.Containers[0].Name)
@@ -561,7 +579,7 @@ func runTopologyManagerTests(f *framework.Framework) {
 
 		configureTopologyManagerInKubelet(f, oldCfg, policy)
 
-		runTopologyManagerNodeAlignmentSuiteTests(f)
+		runTopologyManagerNodeAlignmentSuiteTests(f, numaNodes)
 
 		// restore kubelet config
 		setOldKubeletConfig(f, oldCfg)