Dynamic Port Allocation for GameServers

Implementation of a PortAllocator that ties into the
GameServer controller that will allocate a port on
GameServer creations (if "dynamic" portPolicy is set)
and then also release the port when a GameServer has been
deleted.

This also includes experimental support for node adding,
removal and unscheduling within the cluster.

Closes #14

build/build-image/Dockerfile

@@ -34,7 +34,7 @@ RUN wget -q https://dl.google.com/dl/cloudsdk/release/google-cloud-sdk.zip && un
     /opt/google-cloud-sdk/install.sh --usage-reporting=true --path-update=true --bash-completion=true --rc-path=/root/.bashrc
 
 # update the path for both go and gcloud
-ENV PATH /usr/local/go/bin:/opt/google-cloud-sdk/bin:$PATH
+ENV PATH /usr/local/go/bin:/go/bin:/opt/google-cloud-sdk/bin:$PATH
 
 # RUN gcloud components update
 RUN gcloud components update && gcloud components install kubectl

build/build-image/gen-crd-client.sh

@@ -15,7 +15,8 @@
 # limitations under the License.
 
 rsync -r /go/src/github.com/agonio/agon/vendor/k8s.io/ /go/src/k8s.io/
-/go/src/k8s.io/code-generator/generate-groups.sh "all" \
+cd /go/src/k8s.io/code-generator
+./generate-groups.sh "all" \
     github.com/agonio/agon/pkg/client \
     github.com/agonio/agon/pkg/apis stable:v1alpha1 \
     --go-header-file=/go/src/github.com/agonio/agon/build/boilerplate.go.txt

build/install.yaml

@@ -51,6 +51,10 @@ spec:
           value: "true"
         - name: SIDECAR # overwrite the GameServer sidecar image that is used
           value: ${REGISTRY}/gameservers-sidecar:${VERSION}
+        - name: MIN_PORT
+          value: "7000"
+        - name: MAX_PORT
+          value: "8000"
         livenessProbe:
           httpGet:
             path: /healthz

examples/cpp-simple/gameserver.yaml

@@ -15,11 +15,11 @@
 apiVersion: "stable.agon.io/v1alpha1"
 kind: GameServer
 metadata:
-  name: cpp-simple
+  # generate a unique name
+  # will need to be created with `kubectl create`
+  generateName: cpp-simple-
 spec:
-  portPolicy: "static"
   containerPort: 7654
-  hostPort: 7778
   template:
     spec:
       containers:

examples/gameserver.yaml

@@ -25,21 +25,27 @@
 
 apiVersion: "stable.agon.io/v1alpha1"
 kind: GameServer
+# GameServer Metadata
+# https://v1-8.docs.kubernetes.io/docs/api-reference/v1.8/#objectmeta-v1-meta
 metadata:
-  name: "gds-example"
+  # generateName: "gds-example" # generate a unique name, with the given prefix
+  name: "gds-example" # set a fixed name
 spec:
   # if there is more than one container, specify which one is the game server
   container: example-server
-  # `static` is the only current option. Dynamic port allocated will come in future releases.
-  # When `static` is the policy specified, `hostPort` is required, to specify the port that game clients will connect to
+  # portPolicy has two options:
+  # - "dynamic" (default) the system allocates a free hostPort for the gameserver, for game clients to connect to
+  # - "static", user defines the hostPort that the game client will connect to. Then onus is on the user to ensure that the
+  # port is available. When static is the policy specified, `hostPort` is required to be populated
   portPolicy: "static"
   # the port that is being opened on the game server process
   containerPort: 7654
-  # the port exposed on the host
+  # the port exposed on the host, only required when `portPolicy` is "static". Overwritten when portPolicy is "dynamic".
   hostPort: 7777
   # protocol being used. Defaults to UDP. TCP is the only other option
   protocol: UDP
-  # Pod configuration
+  # Pod template configuration
+  # https://v1-8.docs.kubernetes.io/docs/api-reference/v1.8/#podtemplate-v1-core
   template:
     # pod metadata. Name & Namespace is overwritten
     metadata:

gameservers/controller/controller.go

@@ -15,9 +15,8 @@
 package main
 
 import (
-	"time"
-
 	"net/http"
+	"time"
 
 	"github.com/agonio/agon/pkg/apis/stable"
 	stablev1alpha1 "github.com/agonio/agon/pkg/apis/stable/v1alpha1"
@@ -60,14 +59,16 @@ type Controller struct {
 	gameServerSynced       cache.InformerSynced
 	nodeLister             corelisterv1.NodeLister
 	queue                  workqueue.RateLimitingInterface
+	portAllocator          *PortAllocator
 	server                 *http.Server
 
 	// this allows for overwriting for testing purposes
 	syncHandler func(string) error
 }
 
 // NewController returns a new gameserver crd controller
-func NewController(sidecarImage string,
+func NewController(minPort, maxPort int32,
+	sidecarImage string,
 	alwaysPullSidecarImage bool,
 	kubeClient kubernetes.Interface,
 	kubeInformerFactory informers.SharedInformerFactory,
@@ -89,6 +90,7 @@ func NewController(sidecarImage string,
 		gameServerSynced:       gsInformer.HasSynced,
 		nodeLister:             kubeInformerFactory.Core().V1().Nodes().Lister(),
 		queue:                  workqueue.NewNamedRateLimitingQueue(workqueue.DefaultControllerRateLimiter(), stable.GroupName),
+		portAllocator:          NewPortAllocator(minPort, maxPort, kubeInformerFactory, agonInformerFactory),
 	}
 
 	gsInformer.AddEventHandler(cache.ResourceEventHandlerFuncs{
@@ -118,7 +120,11 @@ func NewController(sidecarImage string,
 
 	mux := http.NewServeMux()
 	mux.HandleFunc("/healthz", func(w http.ResponseWriter, r *http.Request) {
-		w.Write([]byte("ok"))
+		_, err := w.Write([]byte("ok"))
+		if err != nil {
+			logrus.WithError(err).Error("could not send ok response on healthz")
+			w.WriteHeader(http.StatusInternalServerError)
+		}
 	})
 
 	c.server = &http.Server{
@@ -140,7 +146,7 @@ func (c Controller) Run(threadiness int, stop <-chan struct{}) error {
 			logrus.WithError(err).Error("Could not listen on :8080")
 		}
 	}()
-	defer c.server.Close()
+	defer c.server.Close() // nolint: errcheck
 
 	err := c.waitForEstablishedCRD()
 	if err != nil {
@@ -152,6 +158,10 @@ func (c Controller) Run(threadiness int, stop <-chan struct{}) error {
 		return errors.New("failed to wait for caches to sync")
 	}
 
+	if err := c.portAllocator.Run(stop); err != nil {
+		return err
+	}
+
 	logrus.Info("Starting workers...")
 	for i := 0; i < threadiness; i++ {
 		go wait.Until(c.runWorker, time.Second, stop)
@@ -298,6 +308,16 @@ func (c *Controller) syncGameServerBlankState(gs *stablev1alpha1.GameServer) (*s
 	if gs.Status.State == "" && gs.ObjectMeta.DeletionTimestamp.IsZero() {
 		gsCopy := gs.DeepCopy()
 		gsCopy.ApplyDefaults()
+
+		// manage dynamic ports
+		if gsCopy.Spec.PortPolicy == stablev1alpha1.Dynamic {
+			port, err := c.portAllocator.Allocate()
+			if err != nil {
+				return gsCopy, errors.Wrapf(err, "error allocating port for GameServer %s", gsCopy.Name)
+			}
+			gsCopy.Spec.HostPort = port
+		}
+
 		logrus.WithField("gs", gsCopy).Info("Syncing Blank State")
 		gs, err := c.gameServerGetter.GameServers(gs.ObjectMeta.Namespace).Update(gsCopy)
 		return gs, errors.Wrapf(err, "error updating GameServer %s to default values", gs.Name)
@@ -384,6 +404,7 @@ func (c *Controller) syncGameServerRequestReadyState(gs *stablev1alpha1.GameServ
 		gsCopy := gs.DeepCopy()
 		gsCopy.Status.State = stablev1alpha1.Ready
 		gsCopy.Status.Address = addr
+		gsCopy.Status.NodeName = pod.Spec.NodeName
 		// HostPort is always going to be populated, even when dynamic
 		// This will be a double up of information, but it will be easier to read
 		gsCopy.Status.Port = gs.Spec.HostPort
@@ -398,8 +419,8 @@ func (c *Controller) syncGameServerRequestReadyState(gs *stablev1alpha1.GameServ
 func (c *Controller) syncGameServerShutdownState(gs *stablev1alpha1.GameServer) (*stablev1alpha1.GameServer, error) {
 	if gs.Status.State == stablev1alpha1.Shutdown && gs.ObjectMeta.DeletionTimestamp.IsZero() {
 		logrus.WithField("gs", gs).Info("Syncing Shutdown State")
-		// let's be explicit about how we want to shut things down
-		p := metav1.DeletePropagationBackground
+		// Do it in the foreground, so the gameserver gets killed last
+		p := metav1.DeletePropagationForeground
 		err := c.gameServerGetter.GameServers(gs.ObjectMeta.Namespace).Delete(gs.ObjectMeta.Name, &metav1.DeleteOptions{PropagationPolicy: &p})
 		return nil, errors.Wrapf(err, "error deleting Game Server %s", gs.ObjectMeta.Name)
 	}