auto* cc_file = reinterpret_cast<::tensorflow::WritableFile*>(handle);
  TF_SetStatus(status, TF_OK, "");
  ::tensorflow::Set_TF_Status_from_Status(status, cc_file->Sync());
}

void TF_FlushWritableFile(TF_WritableFileHandle* handle, TF_Status* status) {
  auto* cc_file = reinterpret_cast<::tensorflow::WritableFile*>(handle);
  TF_SetStatus(status, TF_OK, "");

  if (fflush(posix_file->handle) != 0)
    TF_SetStatusFromIOError(status, errno, posix_file->filename);
}

static void Sync(const TF_WritableFile* file, TF_Status* status) {
  // For historical reasons, this does the same as `Flush` at the moment.
  // TODO(b/144055243): This should use `fsync`/`sync`.
  Flush(file, status);
}

static void Close(const TF_WritableFile* file, TF_Status* status) {

            gcs_file->object.c_str());
    if (TF_GetCode(status) != TF_OK) return;
    gcs_file->sync_need = false;
  } else {
    TF_SetStatus(status, TF_OK, "");
  }
}

void Sync(const TF_WritableFile* file, TF_Status* status) {
  auto gcs_file = static_cast<GCSFile*>(file->plugin_file);
  TF_VLog(3, "Sync: gs://%s/%s", gcs_file->bucket.c_str(),
          gcs_file->object.c_str());
  Flush(file, status);
}

      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(

  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ops_->flush(file_.get(), plugin_status.get());
  return StatusFromTF_Status(plugin_status.get());
}

Status ModularWritableFile::Sync() {
  if (ops_->sync == nullptr) return Flush();

  UniquePtrTo_TF_Status plugin_status(TF_NewStatus(), TF_DeleteStatus);
  ops_->sync(file_.get(), plugin_status.get());
  return StatusFromTF_Status(plugin_status.get());
}

        handle->TypeString(), " tensor handle.");
    return nullptr;
  }
  tensorflow::Device* device(handle->device());
  if (device != nullptr) {
    status->status = device->Sync();
    if (!status->status.ok()) {
      return nullptr;
    }
  }
  const tensorflow::Tensor* tensor;
  status->status = handle->Tensor(&tensor);
  if (!status->status.ok()) {
    return nullptr;
  }

      // Since the input was forwarded, we released the input handle right away
      // and hence expect the input to be forwarded to the return tensor.
      EXPECT_EQ(orig_ptr, TF_TensorData(t));
    } else {
      // In sync mode, forwarding can't really happen since the client code will
      // have a reference count on the input tensor while the kernel is being
      // executed and thus it cannot be re-used for the return tensor.