Thread-safe access to SubscriptionData

[Yadunund]

This PR introduces SubscriptionData class that manages the lifetime of Zenoh artifacts. It makes access to class members thread-safe where possible. Some z_closure callbacks still work with type erased raw ptrs but this will be addressed when we migrate to zenoh-cpp.

It seems like it is not feasible to store rmw_node_t * as type erased void * in rmw_subscription_t->data like we did with rmw_publisher_t->data .

The problem is that because rmw_subscriptions_t in rmw_wait contains an array of the type erased subscriptions handles (ie, rmw_subscription_t->data ), there is no way to get the SubscriptionData::SharedPtr here

rmw_zenoh/rmw_zenoh_cpp/src/rmw_zenoh.cpp

Lines 3220 to 3221 in cf8e8a2

	auto sub_data =
	static_cast<rmw_zenoh_cpp::rmw_subscription_data_t *>(subscriptions->subscribers[i]);

. ie we cannot traverse the heirarchy from rmw_subscription->data->rmw_node_t->context->impl->get_node_data()->get_sub_data().

The only option we have is storing a rawptr to SubscriptionData in rmw_subscription_t->data but leads to the risk of it becoming a dangling pointer when the SubscriptionData::SharedPtr is deleted.

[clalancette]

This generally looks good. I've left some things to improved, and it needs to be rebased. Once those are done, I'll run some more extensive tests on it.

[clalancette]

Is this a separate bugfix?

[Yadunund]

The approach with capturing the weak_ptr ensures that if the event callback executes for some reason after the last reference to PublisherData::SharedPtr goes out of scope, we would safely exit the callback. This is what I want to do with some of the Zenoh callbacks as well once we switch to zenoh-cpp.
I would have liked to also do it within rmw_subscription_event_init but given that we're storing the raw_ptr to SubscriptionData in rmw_subscription_t->data and hence not able to retrieve the shared_ptr to the same, we are stuck with potential to dereference a dangling pointer if the scenario above. But this scenario should never occur given the implementation of rcl*

[ahcorde]

ros2.repos

rmw_zenoh/yadu/raii-sub

• rmw_zenoh commit: f8ea0c3

ROS Package	Failing test name	Failing test output
rcl	None	None
rcl_action	test_graph__rmw_zenoh_cpp	segfault
rcl_lifecycle	None	None
rcl_yaml_param	None	None
rclcpp	test_intra_process_manager	Failed (Check logs)
rclcpp	test_node_interfaces__node_graph	Failed (Check logs)
rclcpp	test_publisher	Failed (Check logs)
rclcpp	test_events_executor	Timeout (Check logs)
rclcpp	test_wait_set	Failed (Check logs)
rclcpp_action	test_server	Failed (Check logs)
rclcpp_components	None	None
rclcpp_lifecycle	None	-
rmw_zenoh_cpp	None	None
test_cli	None	None
test_cli_remap	test_cli_remapping	TBD

---

There are some new segfaults

• rcl_action -> test_graph__rmw_zenoh_cpp
• rclcpp_Action -> test_server

[clalancette]

This generally looks good to me, and passes my tests locally. We should fix the failing rcl_action and rclcpp_action tests before merging, though.

[Yadunund]

@ahcorde thanks a lot for running those tests ahead of time and reporting the issues. I've fixed them in 2bd7e13.