diff --git a/libglfork.cpp b/libglfork.cpp
index 03f514f..bb42f0d 100644
--- a/libglfork.cpp
+++ b/libglfork.cpp
@@ -259,6 +259,22 @@ static struct PrimusInfo {
   }
 } primus;

+static void cleanup()
+{
+  primus.drawables.clear();
+}
+
+static void register_cleanup_1()
+{
+  atexit(cleanup);
+}
+
+static void register_cleanup()
+{
+  static pthread_once_t once = PTHREAD_ONCE_INIT;
+  pthread_once(&once, register_cleanup_1);
+}
+
 // Thread-specific data
 static __thread struct {
   Display *dpy;
@@ -622,11 +638,6 @@ GLXContext glXCreateContextAttribsARB(Display *dpy, GLXFBConfig config, GLXConte
 void glXDestroyContext(Display *dpy, GLXContext ctx)
 {
   primus.contexts.erase(ctx);
-  // kludge: reap background tasks when deleting the last context
-  // otherwise something will deadlock during unloading the library
-  if (primus.contexts.empty())
-    for (DrawablesInfo::iterator i = primus.drawables.begin(); i != primus.drawables.end(); i++)
-      i->second.reap_workers();
   primus.afns.glXDestroyContext(primus.adpy, ctx);
 }

@@ -720,6 +731,7 @@ void glXSwapBuffers(Display *dpy, GLXDrawable drawable)
     di.actx = ctx;
     di.d.spawn_worker(drawable, display_work);
     di.r.spawn_worker(drawable, readback_work);
+    register_cleanup();
   }
   // Readback thread needs a sync object to avoid reading an incomplete frame
   di.sync = primus.afns.glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);