Ensure compliant subscription API.

[hidmic]

Depends on ros2/rmw#256.

[hidmic]

CI up to test_rmw_implementation and rcl, against all RMW implementations:

• Linux
• Linux-aarch64
• macOS
• Windows

[eboasson]

LGTM. I just realised that this prioritizes the error code from publishing the ros discovery topic (a side effect) over the error code from actually destroying the subscription (the primary operation) if both error, and that might be somewhat counterintuitive. The same is true for destroying the publisher, it's just that I didn't realise it at that time.

That said, in this case and for Cyclone DDS, actually destroying the subscription/publication only fails when its handle is no longer valid, which requires having explicitly deleted the entity, its parent (the DDS subscriber), its grandparent (the DDS participant), greatgrandparent (DDS domain instance) or great-greatgrandparent (the Cyclone DDS library initialisation through the magic DDS_CYCLONEDDS_HANDLE). In all but the last of those cases, the return code will be DDS_RETCODE_BAD_PARAMETER, in the last one, it will be DDS_RETCODE_PRECONDITION_NOT_MET.

Arguably, that makes the possible error codes from trying to publish the update to the ros discovery topic more interesting. I also don't see any reason to suggest changing this now. But as a rule, propagating errors from the primary operation rather than those from side-effects seems to make the most sense.

[hidmic]

I just realised that this prioritizes the error code from publishing the ros discovery topic (a side effect) over the error code from actually destroying the subscription (the primary operation) if both error, and that might be somewhat counterintuitive.

Hmm, that's a very good point. And I'm inclined to change it.

[eboasson]

LGTM!