Merge pull request #63114 from grayluck/maxPodEnv

Automatic merge from submit-queue. If you want to cherry-pick this change to another branch, please follow the instructions <a href="https://github.com/kubernetes/community/blob/master/contributors/devel/cherry-picks.md">here</a>.

Add MAX_PODS env allowing kubelet to be max-pods aware.

**What this PR does / why we need it**:
Add MAX_PODS env allowing kubelet to be max-pods aware. This env will be one of the argument of kubelet. By setting this parameter, correct number of pod capacity for the node will be reported.

**Special notes for your reviewer**:
/assign @MrHohn @jpbetz 

**Release note**:

```release-note
NONE
```

cluster/gce/config-common.sh

@@ -98,6 +98,19 @@ function get-cluster-ip-range {
   echo "${suggested_range}"
 }
 
+# Calculate ip alias range based on max number of pods.
+# Let pow be the smallest integer which is bigger than log2($1 * 2).
+# (32 - pow) will be returned.
+#
+# $1: The number of max pods limitation.
+function get-alias-range-size() {
+  for pow in {0..31}; do
+    if (( 1 << $pow > $1 * 2 )); then
+      echo $((32 - pow))
+      return 0
+    fi
+  done
+}
 # NOTE: Avoid giving nodes empty scopes, because kubelet needs a service account
 # in order to initialize properly.
 NODE_SCOPES="${NODE_SCOPES:-monitoring,logging-write,storage-ro}"

cluster/gce/config-default.sh

@@ -289,8 +289,10 @@ ENABLE_RESCHEDULER="${KUBE_ENABLE_RESCHEDULER:-true}"
 ENABLE_IP_ALIASES=${KUBE_GCE_ENABLE_IP_ALIASES:-false}
 NODE_IPAM_MODE=${KUBE_GCE_NODE_IPAM_MODE:-RangeAllocator}
 if [ ${ENABLE_IP_ALIASES} = true ]; then
-  # Size of ranges allocated to each node. Currently supports only /32 and /24.
+  # Number of Pods that can run on this node.
-  IP_ALIAS_SIZE=${KUBE_GCE_IP_ALIAS_SIZE:-/24}
+  MAX_PODS_PER_NODE=${MAX_PODS_PER_NODE:-110}
+  # Size of ranges allocated to each node.
+  IP_ALIAS_SIZE="/$(get-alias-range-size ${MAX_PODS_PER_NODE})"
   IP_ALIAS_SUBNETWORK=${KUBE_GCE_IP_ALIAS_SUBNETWORK:-${INSTANCE_PREFIX}-subnet-default}
   # If we're using custom network, use the subnet we already create for it as the one for ip-alias.
   # Note that this means SUBNETWORK would override KUBE_GCE_IP_ALIAS_SUBNETWORK in case of custom network.
@@ -305,6 +307,10 @@ if [ ${ENABLE_IP_ALIASES} = true ]; then
   PROVIDER_VARS="${PROVIDER_VARS:-} ENABLE_IP_ALIASES"
   PROVIDER_VARS="${PROVIDER_VARS:-} NODE_IPAM_MODE"
   PROVIDER_VARS="${PROVIDER_VARS:-} SECONDARY_RANGE_NAME"
+elif [[ -n "${MAX_PODS_PER_NODE:-}" ]]; then
+  # Should not have MAX_PODS_PER_NODE set for route-based clusters.
+  echo -e "${color_red}Cannot set MAX_PODS_PER_NODE for route-based projects for ${PROJECT}." >&2
+  exit 1
 fi
 
 # Enable GCE Alpha features.

cluster/gce/config-test.sh

@@ -296,8 +296,10 @@ ENABLE_RESCHEDULER="${KUBE_ENABLE_RESCHEDULER:-true}"
 ENABLE_IP_ALIASES=${KUBE_GCE_ENABLE_IP_ALIASES:-false}
 NODE_IPAM_MODE=${KUBE_GCE_NODE_IPAM_MODE:-RangeAllocator}
 if [ ${ENABLE_IP_ALIASES} = true ]; then
-  # Size of ranges allocated to each node. gcloud current supports only /32 and /24.
+  # Number of Pods that can run on this node.
-  IP_ALIAS_SIZE=${KUBE_GCE_IP_ALIAS_SIZE:-/24}
+  MAX_PODS_PER_NODE=${MAX_PODS_PER_NODE:-110}
+  # Size of ranges allocated to each node.
+  IP_ALIAS_SIZE="/$(get-alias-range-size ${MAX_PODS_PER_NODE})"
   IP_ALIAS_SUBNETWORK=${KUBE_GCE_IP_ALIAS_SUBNETWORK:-${INSTANCE_PREFIX}-subnet-default}
   # If we're using custom network, use the subnet we already create for it as the one for ip-alias.
   # Note that this means SUBNETWORK would override KUBE_GCE_IP_ALIAS_SUBNETWORK in case of custom network.
@@ -312,6 +314,10 @@ if [ ${ENABLE_IP_ALIASES} = true ]; then
   PROVIDER_VARS="${PROVIDER_VARS:-} ENABLE_IP_ALIASES"
   PROVIDER_VARS="${PROVIDER_VARS:-} NODE_IPAM_MODE"
   PROVIDER_VARS="${PROVIDER_VARS:-} SECONDARY_RANGE_NAME"
+elif [[ -n "${MAX_PODS_PER_NODE:-}" ]]; then
+  # Should not have MAX_PODS_PER_NODE set for route-based clusters.
+  echo -e "${color_red}Cannot set MAX_PODS_PER_NODE for route-based projects for ${PROJECT}." >&2
+  exit 1
 fi
 
 # Enable GCE Alpha features.

cluster/gce/util.sh

@@ -677,6 +677,9 @@ function construct-kubelet-flags {
   if [[ -n "${CONTAINER_RUNTIME_ENDPOINT:-}" ]]; then
     flags+=" --container-runtime-endpoint=${CONTAINER_RUNTIME_ENDPOINT}"
   fi
+  if [[ -n "${MAX_PODS_PER_NODE:-}" ]]; then
+    flags+=" --max-pods=${MAX_PODS_PER_NODE}"
+  fi
 
   KUBELET_ARGS="${flags}"
 }
@@ -1136,6 +1139,11 @@ EOF
   if [ -n "${SCHEDULING_ALGORITHM_PROVIDER:-}" ]; then
     cat >>$file <<EOF
 SCHEDULING_ALGORITHM_PROVIDER: $(yaml-quote ${SCHEDULING_ALGORITHM_PROVIDER})
+EOF
+  fi
+  if [ -n "${MAX_PODS_PER_NODE:-}" ]; then
+    cat >>$file <<EOF
+MAX_PODS_PER_NODE: $(yaml-quote ${MAX_PODS_PER_NODE})
 EOF
   fi
 }