remove I/O from signal handler.

[fujitatomoya]

Description

Backport of #2986

I tried to think of the ABI compatible way to backport the #2986, because it is critical issue.
But ended up removing the I/O (including possible rcl_publish) from the signal handler.

• reuse signal_received_ as std::atomic_int to store the signal number. this also changes the memory layout/alighment from 1 byte to 4 bytes. checked with amd64 and arm64 platform, this is also gonna be the breaking ABI change, and not guaranteed by standard.
• external data structure such as code space static variable or per-instance state using instance IDs or pointers, avoiding class member changes. However, this adds complexity and potential thread-safety concerns.

Note

this requires ros2/system_tests#580

Is this user-facing behavior change?

Yes, the user application cannot expect the logging message (or publish) via signal handler.

Did you use Generative AI?

No,

Additional Information

[fujitatomoya]

@jmachowinski do we have other options to backport #2986? i tried to think of the possible approach, but ended up having this PR instead.

[fujitatomoya]

Pulls: #3000, ros2/system_tests#580
Gist: https://gist.githubusercontent.com/fujitatomoya/cb0cb9f2ccbaa3aacf702208a0af568e/raw/8ebcd90eb3c0a40947db66d0fff6f5af5e62ec85/ros2.repos
BUILD args: --packages-above-and-dependencies rclcpp test_rclcpp
TEST args: --packages-above rclcpp test_rclcpp
ROS Distro: kilted
Job: ci_launcher
ci_launcher ran: https://ci.ros2.org/job/ci_launcher/17600

• Linux
• Linux-aarch64
• Linux-rhel
• Windows

[jmachowinski]

I think you can move the info output here