Defensive programming: ACEC is an optimization and not strictly needed

src/core/lib/iomgr/exec_ctx.h

@@ -357,6 +357,11 @@ class ApplicationCallbackExecCtx {
   /** Global shutdown for ApplicationCallbackExecCtx. Called by init. */
   static void GlobalShutdown(void) { gpr_tls_destroy(&callback_exec_ctx_); }
 
+  static bool Available() {
+    return reinterpret_cast<ApplicationCallbackExecCtx*>(
+               gpr_tls_get(&callback_exec_ctx_)) != nullptr;
+  }
+
  private:
   uintptr_t flags_{0u};
   grpc_experimental_completion_queue_functor* head_{nullptr};

src/core/lib/surface/completion_queue.cc

@@ -874,8 +874,15 @@ static void cq_end_op_for_callback(
     cq_finish_shutdown_callback(cq);
   }
 
+  // If possible, schedule the callback onto an existing thread-local
+  // ApplicationCallbackExecCtx, which is a work queue. This is possible for:
+  // 1. The callback is internally-generated and there is an ACEC available
+  // 2. The callback is marked inlineable and there is an ACEC available
+  // 3. We are already running in a background poller thread (which always has
+  //    an ACEC available at the base of the stack).
   auto* functor = static_cast<grpc_experimental_completion_queue_functor*>(tag);
-  if (internal || functor->inlineable ||
+  if (((internal || functor->inlineable) &&
+       grpc_core::ApplicationCallbackExecCtx::Available()) ||
       grpc_iomgr_is_any_background_poller_thread()) {
     grpc_core::ApplicationCallbackExecCtx::Enqueue(functor,
                                                    (error == GRPC_ERROR_NONE));