*: Have keeper notify when shutting down

Avoided race condition in sending dying gasp

Renamed variable

cmd/keeper/cmd/keeper.go

@@ -585,7 +585,7 @@ func (p *PostgresKeeper) usePgrewind(db *cluster.DB) bool {
 	return p.pgSUUsername != "" && p.pgSUPassword != "" && db.Spec.UsePgrewind
 }
 
-func (p *PostgresKeeper) updateKeeperInfo() error {
+func (p *PostgresKeeper) updateKeeperInfo(shuttingDown bool) error {
 	p.localStateMutex.Lock()
 	keeperUID := p.keeperLocalState.UID
 	clusterUID := p.keeperLocalState.ClusterUID
@@ -601,6 +601,13 @@ func (p *PostgresKeeper) updateKeeperInfo() error {
 		log.Warnf("failed to get postgres binary version: %v", err)
 	}
 
+	lastPGState := p.getLastPGState()
+	if shuttingDown {
+		// Tell the sentinels we are shutting this down.
+		// We don't need the mutex since p.getLastPGState()
+		// does a deep copy.
+		lastPGState.Healthy = false
+	}
 	keeperInfo := &cluster.KeeperInfo{
 		InfoUID:    common.UID(),
 		UID:        keeperUID,
@@ -610,7 +617,7 @@ func (p *PostgresKeeper) updateKeeperInfo() error {
 			Maj: maj,
 			Min: min,
 		},
-		PostgresState: p.getLastPGState(),
+		PostgresState: lastPGState,
 
 		CanBeMaster:             p.canBeMaster,
 		CanBeSynchronousReplica: p.canBeSynchronousReplica,
@@ -801,9 +808,9 @@ func (p *PostgresKeeper) getLastPGState() *cluster.PostgresState {
 }
 
 func (p *PostgresKeeper) Start(ctx context.Context) {
-	endSMCh := make(chan struct{})
-	endPgStatecheckerCh := make(chan struct{})
-	endUpdateKeeperInfo := make(chan struct{})
+	endPostgresKeeperSM := make(chan struct{}, 1)
+	endUpdatePGState := make(chan struct{}, 1)
+	endUpdateKeeperInfo := make(chan struct{}, 1)
 
 	var err error
 	var cd *cluster.ClusterData
@@ -828,9 +835,29 @@ func (p *PostgresKeeper) Start(ctx context.Context) {
 
 	_ = p.pgm.StopIfStarted(true)
 
-	smTimerCh := time.NewTimer(0).C
-	updatePGStateTimerCh := time.NewTimer(0).C
-	updateKeeperInfoTimerCh := time.NewTimer(0).C
+	postgresKeeperSMTimer := time.NewTimer(0)
+	updatePGStateTimer := time.NewTimer(0)
+	updateKeeperInfoTimer := time.NewTimer(0)
+
+	// Make sure there is a single goroutine sending KeeperInfo updates.
+	// We want to force an order in calls to updateKeeperInfo, so the
+	// KeeperInfo sent when the context is cancelled is guaranteed to be
+	// the last one.
+	doUpdateKeeperInfo := make(chan struct{}, 1)
+	go func() {
+		defer close(endUpdateKeeperInfo)
+		for range doUpdateKeeperInfo {
+			if err := p.updateKeeperInfo(false); err != nil {
+				log.Errorw("failed to update keeper info", zap.Error(err))
+			}
+			endUpdateKeeperInfo <- struct{}{}
+		}
+		// Once the channel is closed, send a dying gasp
+		if err := p.updateKeeperInfo(true); err != nil {
+			log.Errorw("failed to update keeper info", zap.Error(err))
+		}
+	}()
+
 	for {
 		// The sleepInterval can be updated during normal execution. Ensure we regularly
 		// refresh the metric to account for those changes.
@@ -839,42 +866,40 @@ func (p *PostgresKeeper) Start(ctx context.Context) {
 		select {
 		case <-ctx.Done():
 			log.Debugw("stopping stolon keeper")
+			close(doUpdateKeeperInfo)       // This will notify sentinels we are shutting down
+			for range endUpdateKeeperInfo { // Wait until the dying gasp is sent
+			}
 			if err = p.pgm.StopIfStarted(true); err != nil {
 				log.Errorw("failed to stop pg instance", zap.Error(err))
 			}
 			p.end <- nil
 			return
 
-		case <-smTimerCh:
+		case <-postgresKeeperSMTimer.C:
 			go func() {
 				p.postgresKeeperSM(ctx)
-				endSMCh <- struct{}{}
+				endPostgresKeeperSM <- struct{}{}
 			}()
 
-		case <-endSMCh:
-			smTimerCh = time.NewTimer(p.sleepInterval).C
+		case <-endPostgresKeeperSM:
+			postgresKeeperSMTimer.Reset(p.sleepInterval)
 
-		case <-updatePGStateTimerCh:
-			// updateKeeperInfo two times faster than the sleep interval
+		case <-updatePGStateTimer.C:
+			// updatePGState two times faster than the sleep interval
 			go func() {
 				p.updatePGState(ctx)
-				endPgStatecheckerCh <- struct{}{}
+				endUpdatePGState <- struct{}{}
 			}()
 
-		case <-endPgStatecheckerCh:
-			// updateKeeperInfo two times faster than the sleep interval
-			updatePGStateTimerCh = time.NewTimer(p.sleepInterval / 2).C
+		case <-endUpdatePGState:
+			// updatePGState two times faster than the sleep interval
+			updatePGStateTimer.Reset(p.sleepInterval / 2)
 
-		case <-updateKeeperInfoTimerCh:
-			go func() {
-				if err := p.updateKeeperInfo(); err != nil {
-					log.Errorw("failed to update keeper info", zap.Error(err))
-				}
-				endUpdateKeeperInfo <- struct{}{}
-			}()
+		case <-updateKeeperInfoTimer.C:
+			doUpdateKeeperInfo <- struct{}{}
 
 		case <-endUpdateKeeperInfo:
-			updateKeeperInfoTimerCh = time.NewTimer(p.sleepInterval).C
+			updateKeeperInfoTimer.Reset(p.sleepInterval)
 		}
 	}
 }