Bug report

Required Info:

• Operating System:
  • osrf/ros:jazzy-desktop
• Computer:
  • PC
• ROS2 Version:
  • Jazzy Docker
• Version or commit hash:
  • 996b956
• DDS implementation:
  • Fast DDS

Steps to reproduce issue

When launching an application that is using the lifecycle manager with multiple lifecycle nodes, a deadlock can be triggered.
If during the bringup of the lifecycle nodes, a shutdown is initiated(CTRL+C), the bringup and shutdown sequences of the lifecycle manager will run at the same time.

What then happens is a race condition that could result in the bringup logic deadlocking on spin_until_future_complete of one of the nodes, that will never be completed because shutdown has already been initiated.
Right here:

Expected behavior

Expected behavior is that the bringup sequence will be aborted, and all the nodes shutdown.

On the jazzy branch I have added a stop() function in service_client.hpp. This stop function cancels any running spin* functions on service_client's internal executor.

  void stop()
  {
    if (client_)
    {
      callback_group_executor_.cancel();
    }
  }

If we then call this stop function in the destructor of LifecycleServiceClient (lifecycle_service_client.hpp), the deadlock is avoided.

  ~LifecycleServiceClient()
  {
    change_state_.stop();
  };

I have not tested this on Kilted or Main. I see there have been changes to the lifecycle manager after Jazzy, I think the deadlock is still in there.