Fix /dev/kmsg linking regardless of which image is used

The previous approach to link `/dev/kmsg` to `/dev/console` only works on
images that we build ourselves.

The new approach is to do the linking at runtime, parallel to kind
brining up the cluster:
We run a function parallel to kind, which tries to exec into the control
plane node and do the linking. This should succeed as soon as the node
container is up, in which case the function will terminate gracefully.

As this is fixed in kind already [0], we can remove that hack probably
when kind 0.5.0 is out.

Why is this linking even needed? See also [0]!

[0]: kubernetes-sigs/kind#662

Author: hoegaarden

entrypoint.sh

@@ -188,32 +188,6 @@ docker::stop() {
   log::info "Docker exited after ${duration} seconds."
 }
 
-# Starting with k8s 1.15 nodes need to have /dev/kmsg available, otherwise the
-# kubelet will fail to start.
-# For systems which don't have /dev/kmsg, we try to symlink to /dev/console.
-# This is implemented by injecting a systemd service override file into the
-# image in question, which does the symlinking (if needed) before the kubelet
-# is started.
-kind::hack::inject_kmsg_linking() {
-  local imgName="$1"
-  (
-    tmpDir="$(mktemp -d)"
-    trap 'rm -rf -- "$tmpDir"' EXIT
-
-    cd "$tmpDir"
-    {
-      echo '[Service]'
-      echo 'ExecStartPre=/bin/sh -c "[ -e /dev/kmsg ] || ln -s /dev/console /dev/kmsg"'
-    } > 10-kmsg.conf
-    {
-      echo "FROM ${imgName}"
-      echo "COPY 10-kmsg.conf /etc/systemd/system/kubelet.service.d/10-kmsg.conf"
-    } > Dockerfile
-
-    docker build -t "$imgName" .
-  )
-}
-
 # Until kind/kindnet properly supports non-IPv6 systems, we use flannel as a CNI.
 # To do so, we also need to use a custom kind config when starting kind.
 # See also:
@@ -239,21 +213,46 @@ kind::start::fromSource() {
   # create the node image
   GOPATH="$(pwd)/go" \
     kind build node-image --image "$imageName"
-  kind::hack::inject_kmsg_linking "$imageName"
 
   # bring up kind
+  kind::hack::kmsg_linker "$clusterName" &
   kind create cluster --config "$KIND_CONFIG" --image "$imageName" --name "$clusterName" --loglevel "$loglevel" --retain
 
   # get the (compiled) version of kubectl
   cp ./go/src/k8s.io/kubernetes/_output/dockerized/bin/linux/amd64/kubectl ./bin/kubectl
 }
 
+kind::hack::kmsg_linker() {
+  local node nodeCmd clusterName sleepDur tries
+
+  clusterName="$1"
+  nodeCmd='set -e; [ -e /dev/kmsg ] || ln -s /dev/console /dev/kmsg'
+  sleepDur="1s"
+  tries=300
+
+  log::info 'kmsg-linker starting in the background'
+
+  while (( tries-- ))
+  do
+    sleep "$sleepDur"
+    node="$(kind get nodes --name "$clusterName" 2>/dev/null)" || continue
+    docker exec "$node" sh -c "$nodeCmd" 2>/dev/null || continue
+
+    log::info 'kms-linker was successful, exiting'
+    return
+  done
+
+  log::error "kmsg-linker failed after ${tries} tries"
+  return 1
+}
+
 # Start kind with the (latest) node image published by kind upstream
 kind::start::fromUpstream() {
   local clusterName="$1"
 
   log::warn "no k8s source found, using newest node image from kind upstream"
 
+  kind::hack::kmsg_linker "$clusterName" &
   kind create cluster --config "$KIND_CONFIG" --name "$clusterName" --loglevel "$loglevel" --retain
 
   # get kubectl from upstream