fix race condition in applier-manager

Signed-off-by: Ethan Mosbaugh <[email protected]>

pkg/applier/manager.go

@@ -20,6 +20,7 @@ import (
 	"context"
 	"fmt"
 	"path"
+	"sync"
 	"time"
 
 	"github.com/k0sproject/k0s/internal/pkg/dir"
@@ -39,11 +40,13 @@ type Manager struct {
 	KubeClientFactory kubeutil.ClientFactoryInterface
 
 	// client               kubernetes.Interface
-	applier       Applier
-	bundlePath    string
-	cancelWatcher context.CancelFunc
-	log           *logrus.Entry
-	stacks        map[string]stack
+	applier         Applier
+	bundlePath      string
+	stacks          map[string]stack
+	log             *logrus.Entry
+	startChan       chan struct{}
+	mux             sync.Mutex
+	watcherCancelFn context.CancelFunc
 
 	LeaderElector leaderelector.Interface
 }
@@ -67,35 +70,82 @@ func (m *Manager) Init(ctx context.Context) error {
 
 	m.applier = NewApplier(m.K0sVars.ManifestsDir, m.KubeClientFactory)
 
-	m.LeaderElector.AddAcquiredLeaseCallback(func() {
-		watcherCtx, cancel := context.WithCancel(ctx)
-		m.cancelWatcher = cancel
-		go func() {
-			_ = m.runWatchers(watcherCtx)
-		}()
-	})
-	m.LeaderElector.AddLostLeaseCallback(func() {
-		if m.cancelWatcher != nil {
-			m.cancelWatcher()
-		}
-	})
-
-	return err
+	return nil
 }
 
 // Run runs the Manager
 func (m *Manager) Start(_ context.Context) error {
+	m.log.Debug("Starting")
+	m.startChan = make(chan struct{}, 1)
+
+	m.LeaderElector.AddLostLeaseCallback(m.leaseLost)
+
+	m.LeaderElector.AddAcquiredLeaseCallback(m.leaseAcquired)
+
+	// It's possible that by the time we added the callback, we are already the leader,
+	// If this is true the callback will not be called, so we need to check if we are
+	// the leader and notify the channel manually
+	if m.LeaderElector.IsLeader() {
+		m.leaseAcquired()
+	}
+
+	go m.watchStartChan()
 	return nil
 }
 
+func (m *Manager) watchStartChan() {
+	m.log.Debug("Watching start channel")
+	for range m.startChan {
+		m.log.Info("Acquired leader lease")
+		m.mux.Lock()
+		ctx, cancel := context.WithCancel(context.Background())
+		// If there is a previous cancel func, call it
+		if m.watcherCancelFn != nil {
+			m.watcherCancelFn()
+		}
+		m.watcherCancelFn = cancel
+		m.mux.Unlock()
+		go func() {
+			_ = m.runWatchers(ctx)
+		}()
+	}
+	m.log.Info("Start channel closed, stopping applier-manager")
+}
+
 // Stop stops the Manager
 func (m *Manager) Stop() error {
-	if m.cancelWatcher != nil {
-		m.cancelWatcher()
+	m.log.Info("Stopping applier-manager")
+	// We have no guarantees on concurrency here, so use mutex
+	m.mux.Lock()
+	watcherCancelFn := m.watcherCancelFn
+	m.mux.Unlock()
+	if watcherCancelFn != nil {
+		watcherCancelFn()
 	}
+	close(m.startChan)
+	m.log.Debug("Stopped applier-manager")
 	return nil
 }
 
+func (m *Manager) leaseLost() {
+	m.mux.Lock()
+	defer m.mux.Unlock()
+	m.log.Warn("Lost leader lease, stopping applier-manager")
+
+	watcherCancelFn := m.watcherCancelFn
+	if watcherCancelFn != nil {
+		watcherCancelFn()
+	}
+}
+
+func (m *Manager) leaseAcquired() {
+	m.log.Info("Acquired leader lease")
+	select {
+	case m.startChan <- struct{}{}:
+	default:
+	}
+}
+
 func (m *Manager) runWatchers(ctx context.Context) error {
 	log := logrus.WithField("component", constant.ApplierManagerComponentName)