100% action server CPU usage when cancelling a goal from the client

[jubeira]

Bug report

Required Info:

• Operating System:
  • Ubuntu Bionic 18.04 (as in Nvidia OpenGL Docker image)
• Installation type:
  • Binaries
• Version or commit hash:
  • Crystal pre-release (updated 14-12-2018; version 0.6.1)
• DDS implementation:
  • Checked with FastRTPS and connext_cpp
• Client library (if applicable):
  • rclcpp

Steps to reproduce issue

Start top in a terminal to monitor CPU usage.

Then, from another terminal, run minimal_action_server:

ros2 run examples_rclcpp_minimal_action_server action_server_not_composable

Finally, run minimal_action_client with cancel:

ros2 run examples_rclcpp_minimal_action_client action_client_not_composable_with_cancel

Expected behavior

Action is canceled, everything exits gracefully.

Actual behavior

Action is canceled, but in some cases a CPU thread keeps working at 100% until the server is shut down.

Additional information

This doesn't happen every time the action client is launched, but it's not too hard to reproduce either by launching it a few times. It seems to happen ~50% of the times.
This might be related to ros2/rcl#354.

[sloretz]

I can reproduce this on arm64 using https://ci.ros2.org/view/packaging/job/packaging_linux-aarch64/650/ .

It doesn't appear to be related to cancellation. It also happens with the non-cancel example.

ros2 run examples_rclcpp_minimal_action_client action_client_not_composable

The 100% CPU usage starts as soon as the goal request is sent. It can be caused by calling the service directly

ros2 service call /fibonacci/_action/send_goal example_interfaces/Fibonacci_Goal "{order: 1, action_goal_id: {uuid: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3]}}"