Incorrect signal handling

Bug report

Required Info:

• Operating System:
  • Ubuntu 22.04 & Docker Hub ROS2 Humble images
• Installation type:
  • binaries
• Version or commit hash:
  • Humble
• DDS implementation:
  • Fast-RTPS & Cyclone
• Client library (if applicable):
  • rclcpp & rclpy

Steps to reproduce issue

Start a ROS2 Launch process which spawns at least one node, such as

ros2 launch robot_state_publisher rsp-launch-urdf-inline.xml --noninteractive

Send a SIGTERM to the ROS2 launch process with

kill -TERM $ros2_launch_PID

or

Use the official ROS2 OCI compliant containers from Docker Hub and run with Docker or Podman

podman run --rm --name ros_signal_test ros:humble ros2 launch robot_state_publisher rsp-launch-urdf-inline.xml

And then stop the container with

podman stop ros_signal_test

Expected behavior

ROS2 Launch should catch the SIGTERM and propagate that signal to each of its child nodes.

Actual behavior

SIGTERM is ignored in the first instance, and in the second containerized example SIGTERM is still ignored but all processes are SIGKILL'd after the podman stop timeout.

Additional information

The ROS2 Launch design article says

"if someone attempts to SIGTERM the launch system, they probably did so out of impatience after sending SIGINT to the launch system, and therefore the launch system should attempt to exit quickly."

This is not an accurate assumption.

The Linux signal man page defines, according to the POSIX.1-1990 standard, "SIGINT" as an "Interrupt from keyboard" and "SIGTERM" as a "Termination signal".

The GNU manual entry on termination signals goes further, defining "SIGINT" as being "sent when the user types the INTR character (normally C-c)" and "SIGTERM" as "a generic signal used to cause program termination".

Neither of these sources make reference to any hierarchy or precedence of these two signals.

Users of the ROS2 Launch system are likely to first run into this issue when try to automate the running of their ROS2 nodes on embedded systems, in systemd services, podman/docker containers, or both!

By default systemd services send SIGTERMs to their processes on shutdown, followed by SIGKILLs after a timeout.

Similarly, Podman/Docker containers get sent a SIGTERM to PID1 when running, for example, podman stop ros_signal_test.

The issue with containers cropped up here https://github.com/ros2/launch/issues/473, but it was only resolved for SIGINTs.
An issue of SIGTERMs not being handled was mentioned here https://github.com/ros2/launch/issues/545#issuecomment-1191890187, but I've not been able to find any progress on it.

My current workaround for clean termination of an embedded ROS2 Launch system is to set KillMode=SIGINT for systemd services, or run Podman/Docker containers with the --stop-signal=SIGINT flag or set STOPSIGNAL SIGINT in the Dockerfile.
This only works when ROS2 Launch either detects it's running in non-interactive mode or that is set explicitly with --noninteractive https://github.com/ros2/launch/pull/475.

However, it's oxymoronic to have to both be in "non-interactive" mode and to send a "keyboard interrupt" signal (SIGINT) for clean termination!

---

Feature request

Feature description

Have the ROS2 Launch system catch SIGTERMs and forward the same signal to their child nodes.

Implementation considerations

Windows??

Begin with the ros2 launch reproducer using robot_state_publisher rsp-launch-urdf-inline.xml, then compare direct execution with the Podman or Docker container case. Trace how the launch process handles SIGTERM and verify that termination is forwarded to child nodes without requiring SIGINT; the expected result is clean shutdown before the container timeout.