integ/kubernetes/kubernetes-1.18.1/centos/files/kubelet-cpumanager-infrastructure-pods-use-system-re.patch

From d0e89da9ebcbd9a13051ab5366b6daef2cec9bbe Mon Sep 17 00:00:00 2001
From: Chris Friesen <chris.friesen@windriver.com>
Date: Fri, 27 Sep 2019 14:11:54 -0600
Subject: [PATCH 4/6] kubelet cpumanager infrastructure pods use system
 reserved CPUs

This assigns system infrastructure pods to the "reserved" cpuset
to isolate them from the shared pool of CPUs.

Infrastructure pods include any pods that belong to the kube-system,
armada, cert-manager, vault, platform-deployment-manager, portieris,
notification or flux-helm namespaces.

The implementation is a bit simplistic, it is assumed that the
"reserved" cpuset is large enough to handle all infrastructure pods
CPU allocations.

This also prevents infrastucture pods from using Guaranteed resources.

Signed-off-by: Chris Friesen <chris.friesen@windriver.com>
---
 pkg/kubelet/cm/cpumanager/policy_static.go      | 45 +++++++++++++++++++++++++
 pkg/kubelet/cm/cpumanager/policy_static_test.go | 19 ++++++++++-
 2 files changed, 63 insertions(+), 1 deletion(-)

diff --git a/pkg/kubelet/cm/cpumanager/policy_static.go b/pkg/kubelet/cm/cpumanager/policy_static.go
index e631d5d6a74..e511caf7ab7 100644
--- a/pkg/kubelet/cm/cpumanager/policy_static.go
+++ b/pkg/kubelet/cm/cpumanager/policy_static.go
@@ -32,6 +32,11 @@ import (
 // PolicyStatic is the name of the static policy
 const PolicyStatic policyName = "static"
 
+// Define namespaces used by platform infrastructure pods
+var infraNamespaces = [...]string{
+	"kube-system", "armada", "cert-manager", "platform-deployment-manager", "portieris", "vault", "notification", "flux-helm",
+}
+
 // staticPolicy is a CPU manager policy that does not change CPU
 // assignments for exclusively pinned guaranteed containers after the main
 // container process starts.
@@ -205,6 +210,32 @@ func (p *staticPolicy) assignableCPUs(s state.State) cpuset.CPUSet {
 }
 
 func (p *staticPolicy) Allocate(s state.State, pod *v1.Pod, container *v1.Container) error {
+	// Process infra pods before guaranteed pods
+	if isKubeInfra(pod) {
+		// Container belongs in reserved pool.
+		// We don't want to fall through to the p.guaranteedCPUs() clause below so return either nil or error.
+		if _, ok := s.GetCPUSet(string(pod.UID), container.Name); ok {
+				klog.Infof("[cpumanager] static policy: reserved container already present in state, skipping " +
+					"(namespace: %s, pod UID: %s, pod: %s, container: %s)",
+					pod.Namespace, string(pod.UID), pod.Name, container.Name)
+				return nil
+		}
+
+		cpuset := p.reserved
+		if cpuset.IsEmpty() {
+			// If this happens then someone messed up.
+			return fmt.Errorf("[cpumanager] static policy: reserved container unable to allocate cpus " +
+				"(namespace: %s, pod UID: %s, pod: %s, container: %s); cpuset=%v, reserved:%v",
+				pod.Namespace, string(pod.UID), pod.Name, container.Name, cpuset, p.reserved)
+		}
+		s.SetCPUSet(string(pod.UID), container.Name, cpuset)
+		klog.Infof("[cpumanager] static policy: reserved: AddContainer " +
+			"(namespace: %s, pod UID: %s, pod: %s, container: %s); cpuset=%v",
+			pod.Namespace, string(pod.UID), pod.Name, container.Name, cpuset)
+		return nil
+	}
+
+
 	if numCPUs := p.guaranteedCPUs(pod, container); numCPUs != 0 {
 		klog.Infof("[cpumanager] static policy: Allocate (pod: %s, container: %s)", pod.Name, container.Name)
 		// container belongs in an exclusively allocated pool
@@ -300,6 +331,10 @@ func (p *staticPolicy) guaranteedCPUs(pod *v1.Pod, container *v1.Container) int
 	if cpuQuantity.Value()*1000 != cpuQuantity.MilliValue() {
 		return 0
 	}
+	// Infrastructure pods use reserved CPUs even if they're in the Guaranteed QoS class
+	if isKubeInfra(pod) {
+		return 0
+	}
 	// Safe downcast to do for all systems with < 2.1 billion CPUs.
 	// Per the language spec, `int` is guaranteed to be at least 32 bits wide.
 	// https://golang.org/ref/spec#Numeric_types
@@ -417,3 +452,13 @@ func (p *staticPolicy) generateCPUTopologyHints(availableCPUs cpuset.CPUSet, req
 
 	return hints
 }
+
+// check if a given pod is in a platform infrastructure namespace
+func isKubeInfra(pod *v1.Pod) bool {
+	for _, namespace := range infraNamespaces {
+		if namespace == pod.Namespace {
+			return true
+		}
+	}
+	return false
+}
diff --git a/pkg/kubelet/cm/cpumanager/policy_static_test.go b/pkg/kubelet/cm/cpumanager/policy_static_test.go
index b2982432c13..04947d28055 100644
--- a/pkg/kubelet/cm/cpumanager/policy_static_test.go
+++ b/pkg/kubelet/cm/cpumanager/policy_static_test.go
@@ -747,7 +747,8 @@ func TestStaticPolicyStartWithResvList(t *testing.T) {
 }
 
 func TestStaticPolicyAddWithResvList(t *testing.T) {
-
+	infraPod := makePod("fakePod", "fakeContainer2", "200m", "200m")
+	infraPod.Namespace = "kube-system"
 	testCases := []staticPolicyTestWithResvList{
 		{
 			description:     "GuPodSingleCore, SingleSocketHT, ExpectError",
@@ -789,6 +790,22 @@ func TestStaticPolicyAddWithResvList(t *testing.T) {
 			expCPUAlloc:     true,
 			expCSet:         cpuset.NewCPUSet(4, 5),
 		},
+		{
+			description:     "InfraPod, SingleSocketHT, ExpectAllocReserved",
+			topo:            topoSingleSocketHT,
+			numReservedCPUs: 2,
+			reserved:        cpuset.NewCPUSet(0, 1),
+			stAssignments: state.ContainerCPUAssignments{
+				"fakePod": map[string]cpuset.CPUSet{
+					"fakeContainer100": cpuset.NewCPUSet(2, 3, 6, 7),
+				},
+			},
+			stDefaultCPUSet: cpuset.NewCPUSet(4, 5),
+			pod:             infraPod,
+			expErr:          nil,
+			expCPUAlloc:     true,
+			expCSet:         cpuset.NewCPUSet(0, 1),
+		},
 	}
 
 	testExcl := true
-- 
2.16.6