diff --git a/sdks/go/pkg/beam/log/log.go b/sdks/go/pkg/beam/log/log.go
index feae77b6c971..4c1f5dddb018 100644
--- a/sdks/go/pkg/beam/log/log.go
+++ b/sdks/go/pkg/beam/log/log.go
@@ -22,6 +22,7 @@ import (
 	"context"
 	"fmt"
 	"os"
+	"sync/atomic"
 )
 
 // Severity is the severity of the log message.
@@ -44,9 +45,17 @@ type Logger interface {
 	Log(ctx context.Context, sev Severity, calldepth int, msg string)
 }
 
-var (
-	logger Logger = &Standard{}
-)
+var logger atomic.Value
+
+// concreteLogger works around atomic.Value's requirement that the type
+// be identical for all callers.
+type concreteLogger struct {
+	Logger
+}
+
+func init() {
+	logger.Store(&concreteLogger{&Standard{}})
+}
 
 // SetLogger sets the global Logger. Intended to be called during initialization
 // only.
@@ -54,13 +63,13 @@ func SetLogger(l Logger) {
 	if l == nil {
 		panic("Logger cannot be nil")
 	}
-	logger = l
+	logger.Store(&concreteLogger{l})
 }
 
 // Output logs the given message to the global logger. Calldepth is the count
 // of the number of frames to skip when computing the file name and line number.
 func Output(ctx context.Context, sev Severity, calldepth int, msg string) {
-	logger.Log(ctx, sev, calldepth+1, msg) // +1 for this frame
+	logger.Load().(Logger).Log(ctx, sev, calldepth+1, msg) // +1 for this frame
 }
 
 // User-facing logging functions.
diff --git a/sdks/go/pkg/beam/runners/universal/extworker/extworker.go b/sdks/go/pkg/beam/runners/universal/extworker/extworker.go
index dc75c7c8ca5b..ffc8f8e47c09 100644
--- a/sdks/go/pkg/beam/runners/universal/extworker/extworker.go
+++ b/sdks/go/pkg/beam/runners/universal/extworker/extworker.go
@@ -65,6 +65,12 @@ func (s *Loopback) StartWorker(ctx context.Context, req *fnpb.StartWorkerRequest
 	log.Infof(ctx, "starting worker %v", req.GetWorkerId())
 	s.mu.Lock()
 	defer s.mu.Unlock()
+	if s.workers == nil {
+		return &fnpb.StartWorkerResponse{
+			Error: "worker pool shutting down",
+		}, nil
+	}
+
 	if _, ok := s.workers[req.GetWorkerId()]; ok {
 		return &fnpb.StartWorkerResponse{
 			Error: fmt.Sprintf("worker with ID %q already exists", req.GetWorkerId()),
@@ -92,6 +98,10 @@ func (s *Loopback) StopWorker(ctx context.Context, req *fnpb.StopWorkerRequest)
 	log.Infof(ctx, "stopping worker %v", req.GetWorkerId())
 	s.mu.Lock()
 	defer s.mu.Unlock()
+	if s.workers == nil {
+		// Worker pool is already shutting down, so no action is needed.
+		return &fnpb.StopWorkerResponse{}, nil
+	}
 	if cancelfn, ok := s.workers[req.GetWorkerId()]; ok {
 		cancelfn()
 		delete(s.workers, req.GetWorkerId())
@@ -106,12 +116,15 @@ func (s *Loopback) StopWorker(ctx context.Context, req *fnpb.StopWorkerRequest)
 // Stop terminates the service and stops all workers.
 func (s *Loopback) Stop(ctx context.Context) error {
 	s.mu.Lock()
-	defer s.mu.Unlock()
 
 	log.Infof(ctx, "stopping Loopback, and %d workers", len(s.workers))
-	s.workers = map[string]context.CancelFunc{}
-	s.lis.Close()
+	s.workers = nil
 	s.rootCancel()
+
+	// There can be a deadlock between the StopWorker RPC and GracefulStop
+	// which waits for all RPCs to finish, so it must be outside the critical section.
+	s.mu.Unlock()
+
 	s.grpcServer.GracefulStop()
 	return nil
 }