k8s.io/kubernetes@v1.29.3/test/e2e_node/oomkiller_linux_test.go (about) 1 /* 2 Copyright 2022 The Kubernetes Authors. 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 e2enode 18 19 import ( 20 "context" 21 "fmt" 22 23 "github.com/onsi/gomega" 24 v1 "k8s.io/api/core/v1" 25 "k8s.io/apimachinery/pkg/api/resource" 26 metav1 "k8s.io/apimachinery/pkg/apis/meta/v1" 27 utilfeature "k8s.io/apiserver/pkg/util/feature" 28 "k8s.io/kubernetes/pkg/features" 29 kubeletconfig "k8s.io/kubernetes/pkg/kubelet/apis/config" 30 "k8s.io/kubernetes/test/e2e/framework" 31 e2epod "k8s.io/kubernetes/test/e2e/framework/pod" 32 admissionapi "k8s.io/pod-security-admission/api" 33 34 "github.com/onsi/ginkgo/v2" 35 libcontainercgroups "github.com/opencontainers/runc/libcontainer/cgroups" 36 ) 37 38 type testCase struct { 39 name string 40 podSpec *v1.Pod 41 oomTargetContainerName string 42 } 43 44 // KubeReservedMemory is default fraction value of node capacity memory to 45 // be reserved for K8s components. 46 const KubeReservedMemory = 0.35 47 48 var _ = SIGDescribe("OOMKiller for pod using more memory than node allocatable [LinuxOnly]", framework.WithSerial(), func() { 49 f := framework.NewDefaultFramework("nodeallocatable-oomkiller-test") 50 f.NamespacePodSecurityEnforceLevel = admissionapi.LevelPrivileged 51 52 testCases := []testCase{ 53 { 54 name: "single process container without resource limits", 55 oomTargetContainerName: "oomkill-nodeallocatable-container", 56 podSpec: getOOMTargetPod("oomkill-nodeallocatable-pod", "oomkill-nodeallocatable-container", 57 getOOMTargetContainerWithoutLimit), 58 }, 59 } 60 61 for _, testCase := range testCases { 62 runOomKillerTest(f, testCase, KubeReservedMemory) 63 } 64 }) 65 66 var _ = SIGDescribe("OOMKiller [LinuxOnly]", framework.WithNodeConformance(), func() { 67 f := framework.NewDefaultFramework("oomkiller-test") 68 f.NamespacePodSecurityLevel = admissionapi.LevelPrivileged 69 70 testCases := []testCase{ 71 { 72 name: "single process container", 73 oomTargetContainerName: "oomkill-single-target-container", 74 podSpec: getOOMTargetPod("oomkill-target-pod", "oomkill-single-target-container", 75 getOOMTargetContainer), 76 }, 77 { 78 name: "init container", 79 oomTargetContainerName: "oomkill-target-init-container", 80 podSpec: getInitContainerOOMTargetPod("initcontinar-oomkill-target-pod", "oomkill-target-init-container", 81 getOOMTargetContainer), 82 }, 83 } 84 85 // If using cgroup v2, we set memory.oom.group=1 for the container cgroup so that any process which gets OOM killed 86 // in the process, causes all processes in the container to get OOM killed 87 if libcontainercgroups.IsCgroup2UnifiedMode() { 88 testCases = append(testCases, testCase{ 89 name: "multi process container", 90 oomTargetContainerName: "oomkill-multi-target-container", 91 podSpec: getOOMTargetPod("oomkill-target-pod", "oomkill-multi-target-container", 92 getOOMTargetContainerMultiProcess), 93 }) 94 } 95 for _, tc := range testCases { 96 runOomKillerTest(f, tc, 0) 97 } 98 }) 99 100 func runOomKillerTest(f *framework.Framework, testCase testCase, kubeReservedMemory float64) { 101 ginkgo.Context(testCase.name, func() { 102 // Update KubeReservedMemory in KubeletConfig. 103 if kubeReservedMemory > 0 { 104 tempSetCurrentKubeletConfig(f, func(ctx context.Context, initialConfig *kubeletconfig.KubeletConfiguration) { 105 if initialConfig.KubeReserved == nil { 106 initialConfig.KubeReserved = map[string]string{} 107 } 108 // There's a race condition observed between system OOM and cgroup OOM if node alocatable 109 // memory is equal to node capacity. Hence, reserving a fraction of node's memory capacity for 110 // K8s components such that node allocatable memory is less than node capacity to 111 // observe OOM kills at cgroup level instead of system OOM kills. 112 initialConfig.KubeReserved["memory"] = fmt.Sprintf("%d", int(kubeReservedMemory*getLocalNode(context.TODO(), f).Status.Capacity.Memory().AsApproximateFloat64())) 113 }) 114 } 115 116 ginkgo.BeforeEach(func() { 117 ginkgo.By("setting up the pod to be used in the test") 118 e2epod.NewPodClient(f).Create(context.TODO(), testCase.podSpec) 119 }) 120 121 ginkgo.It("The containers terminated by OOM killer should have the reason set to OOMKilled", func() { 122 cfg, configErr := getCurrentKubeletConfig(context.TODO()) 123 framework.ExpectNoError(configErr) 124 if utilfeature.DefaultFeatureGate.Enabled(features.NodeSwap) { 125 // If Swap is enabled, we should test OOM with LimitedSwap. 126 // UnlimitedSwap allows for workloads to use unbounded swap which 127 // makes testing OOM challenging. 128 // We are not able to change the default for these conformance tests, 129 // so we will skip these tests if swap is enabled. 130 if cfg.MemorySwap.SwapBehavior == "" || cfg.MemorySwap.SwapBehavior == "UnlimitedSwap" { 131 ginkgo.Skip("OOMKiller should not run with UnlimitedSwap") 132 } 133 } 134 ginkgo.By("Waiting for the pod to be failed") 135 err := e2epod.WaitForPodTerminatedInNamespace(context.TODO(), f.ClientSet, testCase.podSpec.Name, "", f.Namespace.Name) 136 framework.ExpectNoError(err, "Failed waiting for pod to terminate, %s/%s", f.Namespace.Name, testCase.podSpec.Name) 137 138 ginkgo.By("Fetching the latest pod status") 139 pod, err := f.ClientSet.CoreV1().Pods(f.Namespace.Name).Get(context.TODO(), testCase.podSpec.Name, metav1.GetOptions{}) 140 framework.ExpectNoError(err, "Failed to get the recent pod object for name: %q", pod.Name) 141 142 ginkgo.By("Verifying the OOM target container has the expected reason") 143 verifyReasonForOOMKilledContainer(pod, testCase.oomTargetContainerName) 144 }) 145 146 ginkgo.AfterEach(func() { 147 ginkgo.By(fmt.Sprintf("deleting pod: %s", testCase.podSpec.Name)) 148 e2epod.NewPodClient(f).DeleteSync(context.TODO(), testCase.podSpec.Name, metav1.DeleteOptions{}, framework.PodDeleteTimeout) 149 }) 150 }) 151 } 152 153 func verifyReasonForOOMKilledContainer(pod *v1.Pod, oomTargetContainerName string) { 154 container := e2epod.FindContainerStatusInPod(pod, oomTargetContainerName) 155 if container == nil { 156 framework.Failf("OOM target pod %q, container %q does not have the expected state terminated", pod.Name, oomTargetContainerName) 157 } 158 if container.State.Terminated == nil { 159 framework.Failf("OOM target pod %q, container %q is not in the terminated state", pod.Name, container.Name) 160 } 161 gomega.Expect(container.State.Terminated.ExitCode).To(gomega.Equal(int32(137)), 162 "pod: %q, container: %q has unexpected exitCode: %q", pod.Name, container.Name, container.State.Terminated.ExitCode) 163 164 // This check is currently causing tests to flake on containerd & crio, https://github.com/kubernetes/kubernetes/issues/119600 165 // so we'll skip the reason check if we know its going to fail. 166 // TODO: Remove this once https://github.com/containerd/containerd/issues/8893 is resolved 167 if container.State.Terminated.Reason == "OOMKilled" { 168 gomega.Expect(container.State.Terminated.Reason).To(gomega.Equal("OOMKilled"), 169 "pod: %q, container: %q has unexpected reason: %q", pod.Name, container.Name, container.State.Terminated.Reason) 170 } 171 172 } 173 174 func getOOMTargetPod(podName string, ctnName string, createContainer func(name string) v1.Container) *v1.Pod { 175 return &v1.Pod{ 176 ObjectMeta: metav1.ObjectMeta{ 177 Name: podName, 178 }, 179 Spec: v1.PodSpec{ 180 RestartPolicy: v1.RestartPolicyNever, 181 Containers: []v1.Container{ 182 createContainer(ctnName), 183 }, 184 }, 185 } 186 } 187 188 func getInitContainerOOMTargetPod(podName string, ctnName string, createContainer func(name string) v1.Container) *v1.Pod { 189 return &v1.Pod{ 190 ObjectMeta: metav1.ObjectMeta{ 191 Name: podName, 192 }, 193 Spec: v1.PodSpec{ 194 RestartPolicy: v1.RestartPolicyNever, 195 InitContainers: []v1.Container{ 196 createContainer(ctnName), 197 }, 198 Containers: []v1.Container{ 199 { 200 Name: "busybox", 201 Image: busyboxImage, 202 }, 203 }, 204 }, 205 } 206 } 207 208 // getOOMTargetContainer returns a container with a single process, which attempts to allocate more memory than is 209 // allowed by the container memory limit. 210 func getOOMTargetContainer(name string) v1.Container { 211 return v1.Container{ 212 Name: name, 213 Image: busyboxImage, 214 Command: []string{ 215 "sh", 216 "-c", 217 // use the dd tool to attempt to allocate 20M in a block which exceeds the limit 218 "sleep 5 && dd if=/dev/zero of=/dev/null bs=20M", 219 }, 220 Resources: v1.ResourceRequirements{ 221 Requests: v1.ResourceList{ 222 v1.ResourceMemory: resource.MustParse("15Mi"), 223 }, 224 Limits: v1.ResourceList{ 225 v1.ResourceMemory: resource.MustParse("15Mi"), 226 }, 227 }, 228 } 229 } 230 231 // getOOMTargetContainerMultiProcess returns a container with two processes, one of which attempts to allocate more 232 // memory than is allowed by the container memory limit, and a second process which just sleeps. 233 func getOOMTargetContainerMultiProcess(name string) v1.Container { 234 return v1.Container{ 235 Name: name, 236 Image: busyboxImage, 237 Command: []string{ 238 "sh", 239 "-c", 240 // use the dd tool to attempt to allocate 20M in a block which exceeds the limit 241 "sleep 5 && dd if=/dev/zero of=/dev/null bs=20M & sleep 86400", 242 }, 243 Resources: v1.ResourceRequirements{ 244 Requests: v1.ResourceList{ 245 v1.ResourceMemory: resource.MustParse("15Mi"), 246 }, 247 Limits: v1.ResourceList{ 248 v1.ResourceMemory: resource.MustParse("15Mi"), 249 }, 250 }, 251 } 252 } 253 254 // getOOMTargetContainerWithoutLimit returns a container with a single process which attempts to allocate more memory 255 // than node allocatable and doesn't have resource limits set. 256 func getOOMTargetContainerWithoutLimit(name string) v1.Container { 257 return v1.Container{ 258 Name: name, 259 Image: busyboxImage, 260 Command: []string{ 261 "sh", 262 "-c", 263 // use the dd tool to attempt to allocate huge block of memory which exceeds the node allocatable 264 "sleep 5 && dd if=/dev/zero of=/dev/null iflag=fullblock count=10 bs=10G", 265 }, 266 } 267 }