Merge pull request kubernetes#76505 from rosti/fix-join-cri-detection

kubeadm: Don't error out on join with --cri-socket override

Author: k8s-ci-robot

cmd/kubeadm/app/util/config/cluster.go

@@ -93,6 +93,15 @@ func getInitConfigurationFromCluster(kubeconfigDir string, client clientset.Inte
 		if err := getAPIEndpoint(configMap.Data, initcfg.NodeRegistration.Name, &initcfg.LocalAPIEndpoint); err != nil {
 			return nil, errors.Wrap(err, "failed to getAPIEndpoint")
 		}
+	} else {
+		// In the case where newControlPlane is true we don't go through getNodeRegistration() and initcfg.NodeRegistration.CRISocket is empty.
+		// This forces DetectCRISocket() to be called later on, and if there is more than one CRI installed on the system, it will error out,
+		// while asking for the user to provide an override for the CRI socket. Even if the user provides an override, the call to
+		// DetectCRISocket() can happen too early and thus ignore it (while still erroring out).
+		// However, if newControlPlane == true, initcfg.NodeRegistration is not used at all and it's overwritten later on.
+		// Thus it's necessary to supply some default value, that will avoid the call to DetectCRISocket() and as
+		// initcfg.NodeRegistration is discarded, setting whatever value here is harmless.
+		initcfg.NodeRegistration.CRISocket = constants.DefaultDockerCRISocket
 	}
 	return initcfg, nil
 }