Add a test for destroy_node while spinning

[cottsay]

I'm not sure if we ever intended to support calling Node.destroy_node() while another thread is spinning or not, but if you do so, the spinner will die with a segmentation fault.

This PR doesn't actually address the issue, it just adds a test to demonstrate the problem.

With rmw_cyclonedds_cpp:

• Linux
• Linux-aarch64
• macOS
• Windows

With rmw_fastrtps_cpp:

• Linux
• Linux-aarch64
• macOS
• Windows

[cottsay]

After some digging, it appears that the exception generated by rmw_fastrtps_cpp is expected.

The segfault when using rmw_cyclonedds_cpp results from the QoS event handlers being referenced during the call to rmw_wait after it is awoken. The call to destroy_node is deallocating those event handlers prior to rmw_wait waking up, hence the segfault.

Here is an example of this occuring on a macOS build: #667 (comment)

This doesn't appear to be a problem in rmw_fastrtps_cpp because it doesn't seem reference the data associated with the event handlers after rmw_wait wakes up in the same way that rmw_cyclonedds_cpp does.

The rmw_wait event handling was added to rmw_cyclonedds_cpp in ros2/rmw_cyclonedds#88, and the site of the fault is here:
https://github.com/ros2/rmw_cyclonedds/blob/d024823043504ea40af24bf22365a21cd203df55/rmw_cyclonedds_cpp/src/rmw_node.cpp#L3291

[clalancette]

As we discussed, I've converted this to a draft PR and assigned it to one of the rclpy maintainers. This is probably something that we should try to fix for Galactic, so I'm also putting it on the Galactic board.

[clalancette]

Now that we've merged #761, I think it is probably a good idea to rebase this one and see if it is fixed.

[clalancette]

CI:

• Linux
• Linux-aarch64
• macOS
• Windows

[cottsay]

Evidently I need to review my knowledge of the English alphabet.

After fixing the import order:

• Linux
• Linux-aarch64
• macOS
• Windows

[clalancette]

Seems like a reasonable test to have to me! Should catch these kinds of issues in the future.