From f6f872a8fcee04890383065d3683d6251b35e160 Mon Sep 17 00:00:00 2001
From: Jeremy Peterson <jerpeter@redhat.com>
Date: Wed, 8 Mar 2023 23:41:21 -0500
Subject: [PATCH] Support Graceful Shutdown

If the kubelet where the the sriov pod is running has gracefulShutdown
configured, we'll delay in preStop for a while if and while /tmp/sriov-delay-shutdown
exists, up to a maximum wait of 10 minutes (less than the 15 minutes which is specified in
the daemonset pod's terminationGracePeriodSeconds).
---
 bindata/manifests/daemon/daemonset.yaml |  1 +
 bindata/scripts/clean-k8s-services.sh   | 28 ++++++++++++++--
 pkg/daemon/daemon.go                    | 44 ++++++++++++++++++++-----
 3 files changed, 63 insertions(+), 10 deletions(-)

diff --git a/bindata/manifests/daemon/daemonset.yaml b/bindata/manifests/daemon/daemonset.yaml
index 2a5d5bdaf..b10879fa2 100644
--- a/bindata/manifests/daemon/daemonset.yaml
+++ b/bindata/manifests/daemon/daemonset.yaml
@@ -93,6 +93,7 @@ spec:
         volumeMounts:
         - name: cnibin
           mountPath: /host/opt/cni/bin
+      terminationGracePeriodSeconds: 900
       volumes:
       - name: host
         hostPath:
diff --git a/bindata/scripts/clean-k8s-services.sh b/bindata/scripts/clean-k8s-services.sh
index 59cbc5fd7..737a5d256 100755
--- a/bindata/scripts/clean-k8s-services.sh
+++ b/bindata/scripts/clean-k8s-services.sh
@@ -1,12 +1,36 @@
 #!/bin/bash
 
+chroot_path="/proc/1/root"
+delay_shutdown_path="$chroot_path/tmp/sriov-delay-shutdown"
+kubelet_config_path="$chroot_path/etc/kubernetes/kubelet.conf"
+
+# 10 minutes - this should be shorter than the time that is specifed for the
+# terminationGracePeriodSeconds in the daemonset's pod spec, so that everything
+# else in the preStop hook has time to run and the Pod can be terminated properly.
+wait_time=600
+
+# If the kubelet is configured to shutdown gracefully (>0s shutdownGracePeriod), we need to wait for
+# things to settle before shutting down the node.
+if [ -f "$delay_shutdown_path" ]; then
+  if grep "$kubelet_config_path" -e shutdownGracePeriod | grep -qv \"0s\"; then
+    start=$(date +%s)
+    touch "$chroot_path/var/log/sriov-delay-start"
+    while [ $(( $(date +%s) - $start )) -lt $wait_time ]; do
+      if [ ! -f "$delay_shutdown_path" ]; then  # don't have to wait anymore
+        break
+      fi
+      sleep 1
+    done
+    rm -f "$delay_shutdown_path"
+    touch "$chroot_path/var/log/sriov-delay-end"
+  fi
+fi
+
 if [ "$CLUSTER_TYPE" == "openshift" ]; then
   echo "openshift cluster"
   exit
 fi
 
-chroot_path="/host"
-
 function clean_services() {
   # Remove switchdev service files
   rm -f $chroot_path/etc/systemd/system/switchdev-configuration-after-nm.service
diff --git a/pkg/daemon/daemon.go b/pkg/daemon/daemon.go
index 727ffe642..c6459f43e 100644
--- a/pkg/daemon/daemon.go
+++ b/pkg/daemon/daemon.go
@@ -52,6 +52,9 @@ const (
 	// maxUpdateBackoff is the maximum time to react to a change as we back off
 	// in the face of errors.
 	maxUpdateBackoff = 60 * time.Second
+
+	// the presence of this file indicates that the sriov shutdown should be delayed
+	delayShutdownPath = "/host/tmp/sriov-delay-shutdown"
 )
 
 type Message struct {
@@ -612,6 +615,22 @@ func (dn *Daemon) completeDrain() error {
 		glog.Errorf("completeDrain(): failed to annotate node: %v", err)
 		return err
 	}
+
+	if _, err := os.Stat(delayShutdownPath); err != nil {
+		if os.IsNotExist(err) {
+			// delayShutdownPath does not exist, so we don't need to do anything
+			return nil
+		}
+
+		glog.Errorf("completeDrain(): error checking file status %v: %v", delayShutdownPath, err)
+		return err
+	}
+
+	if err := os.Remove(delayShutdownPath); err != nil {
+		glog.Errorf("completeDrain(): failed to remove file %v: %v", delayShutdownPath, err)
+		return err
+	}
+
 	return nil
 }
 
@@ -679,15 +698,16 @@ func rebootNode() {
 		glog.Errorf("rebootNode(): %v", err)
 	}
 	defer exit()
-	// creates a new transient systemd unit to reboot the system.
-	// We explictily try to stop kubelet.service first, before anything else; this
-	// way we ensure the rest of system stays running, because kubelet may need
-	// to do "graceful" shutdown by e.g. de-registering with a load balancer.
-	// However note we use `;` instead of `&&` so we keep rebooting even
-	// if kubelet failed to shutdown - that way the machine will still eventually reboot
-	// as systemd will time out the stop invocation.
+	// creates a new transient systemd unit to reboot the system that
+	// reboots the system using `systemctl reboot``
+	// by shutting down the system this way instead via `reboot`,
+	// when kubelet is configured with a shutdownGracePeriod, then it will
+	// be give some time to pods to run their preStop scripts and respond to
+	// SIGTERM by terminating gracefully before being forcefully killed via
+	// SIGKILL. Stopping the kubelet service and then immediately running
+	// `reboot` just results in all pods being immediately killed
 	cmd := exec.Command("systemd-run", "--unit", "sriov-network-config-daemon-reboot",
-		"--description", "sriov-network-config-daemon reboot node", "/bin/sh", "-c", "systemctl stop kubelet.service; reboot")
+		"--description", "sriov-network-config-daemon reboot node", "/bin/sh", "-c", "systemctl reboot")
 
 	if err := cmd.Run(); err != nil {
 		glog.Errorf("failed to reboot node: %v", err)
@@ -933,6 +953,14 @@ func (dn *Daemon) drainNode() error {
 		return err
 	}
 	glog.Info("drainNode(): drain complete")
+
+	file, err := os.Create(delayShutdownPath)
+	if err != nil {
+		glog.Errorf("drainNode(): failed to create file %v %v", delayShutdownPath, err)
+		return err
+	}
+	defer file.Close()
+
 	return nil
 }