[iibaranov-IG]

Fixes #422

When events_ws_port and control_protocol_ws_port matched, make_node_server rejected the configuration before creating its listeners. The WebSocket endpoints are already route-distinct: IS-07 uses /x-nmos/events/... and IS-12 uses /x-nmos/ncp/v1.0.

This change installs one listener for a shared port and dispatches handshake, open, close, and message callbacks by endpoint path. IS-07 and IS-12 retain separate subscription/connection state and message threads. Existing configurations with distinct ports continue to create independent listeners. IS-12 validation now accepts only configured /x-nmos/ncp/<version> routes, so invalid or conflicting paths are rejected instead of being claimed by the control endpoint.

Tests cover same-port, separate-port, IS-07-only, and valid/invalid/conflicting IS-12 routes.

Validation: Debug build and all 189 CTest tests pass on macOS/Apple Silicon.

[lo-simon]

Hi @iibaranov-IG , thank you so much for your contribution. I have refined your suggestions to make them more general and future-proof.

[garethsb]

Why deque?

[lo-simon]

This is because make_ws_api_listener() captures the handlers by reference; each websocket_listener ends up storing a std::reference_wrapper pointing directly at the memory address of a specific websocket_listener_handlers object living inside combined_ws_handlers.

If combined_ws_handlers were a std::vector, calling .push_back() in a loop (once per WebSocket port being configured) can trigger reallocation when the vector's capacity is exceeded. Reallocation copies/moves all existing elements to a new, larger memory block and frees the old one. That invalidates every reference/pointer/iterator to previously inserted elements; see example below:

e.g. if with a std::vector

for (auto& ws_handler : node_server.ws_handlers)   // e.g. 2 iterations: events port, control protocol port
{
    ...
    node_server.combined_ws_handlers.push_back(nmos::make_ws_api_router(path_handlers));
    // if this push_back reallocates, any websocket_listener already
    // constructed in a PREVIOUS iteration (which holds std::ref to an
    // earlier combined_ws_handlers element) is now dangling!
    node_server.ws_listeners.insert({ ws_handler.first, nmos::make_ws_api_listener(..., node_server.combined_ws_handlers.back(), ...) });
}

But std::deque guarantees that inserting at the front or back never invalidates references or pointers to other elements (only iterators may be invalidated, but we're not holding onto iterators - we're holding onto references inside std::ref, which deque preserves across push_back). Internally, deque is implemented as a sequence of fixed-size chunks, so growing it just adds a new chunk rather than reallocating/moving all existing elements.

This matches exactly what we need here: every websocket_listener_handlers added in earlier loop iterations must remain at a stable address for the entire lifetime of node_server, even as more are added in later iterations.

[garethsb]

Makes sense, thanks!

std::map<host_port, websocket_listener_handlers> would keep references stable too, and would be keyed the same way as ws_listeners, so each listener's handlers are easy to find. Or we could keep each listener and its handlers together as a pair value in one map? Or maybe no need.

[garethsb]

In settings.cpp, the http_port defaulting still says "can't share a port between the events ws and the control protocol ws" and sets ncp_ws_port = ws_port + 1. The PR should fix the comment or maybe even change to share by default?

[garethsb]

Needs at least a read lock (and might as well use the same simple versions name), like:

nmos-cpp/Development/nmos/node_api.cpp

Line 41 in b6c8367

const auto versions = with_read_lock(model.mutex, [&model] { return nmos::is04_versions::from_settings(model.settings); });

[garethsb]

would it make sense to capture find_handlers via a shared_ptr to avoid making four copies? not sure