[stewartadam]

Problem

When a data packet names a synchronisation address S, the receiver called listen_universes(&[S]). Joining S's multicast group is right, because the sync packets arrive there. But listen_universes also adds S to the universes the user asked for, which had three side effects:

• data packets sent on universe S were passed to the application even though it never listened to S;
• the group was never left, even after sources stopped using S;
• memberships piled up. Linux caps multicast memberships per socket (igmp_max_memberships, 20 by default), so a network that cycles through many sync addresses eventually makes the join fail, and that error came back from recv for every such packet.

Change

• New sync_universes: HashMap<u16, Instant> on SacnReceiver. It records each sync address named by a data packet and when it was last named.
• handle_data_packet calls a new track_sync_universe. That joins the group (if it isn't already joined, multicast is enabled, and the user isn't already listening to S) and refreshes the timestamp. S is not added to the data universes.
• handle_sync_packet accepts sync packets for universes the user listens to or that are in sync_universes.
• recv now calls expire_sync_universes next to the existing waiting-data timeout check. It drops sync universes that haven't been named for E131_NETWORK_DATA_LOSS_TIMEOUT, and leaves their group unless the user listens to that universe.
• listen_universes doesn't join a group already joined for sync, since a second join fails with EADDRINUSE on Linux. mute_universe doesn't leave a group still needed for sync.
• Drop also leaves the groups joined only for sync.

Behaviour change: data sent on a universe that is only used as a sync address is no longer delivered unless the application listens to that universe.

Tests

Three unit tests in receive.rs:

• the sync universe isn't added to the data universes, data on it is ignored, and the sync packet still releases the waiting data;
• an expired sync universe is forgotten (its sync packets are ignored) and can be tracked again;
• listening to and then muting a universe that's also used for sync neither joins twice nor leaves early. I checked this test fails with AddrInUse without the listen_universes guard. Its last assertion relies on leaving an unjoined group erroring, so it's skipped on Windows.

cargo test --lib passes. The #[ignore]d sync integration tests in tests/ipv4_tests.rs need specific LAN addresses, so I ran them with those addresses switched to loopback (not committed). All 11 pass.