integration_tests::service::test_service_ipv4_unknown_address flake

[rcgoodfellow]

I had integration_tests::service::test_service_ipv4_unknown_address fail in CI. Run is here. Rerunning without changes was sufficient to have the test pass.

[Nieuwejaar]

I've seen this a few times in CI myself, but I've just run 200+ iterations locally without hitting it. There should be nothing non-deterministic or timing-sensitive about it, which is the first place I look for this sort of issue.

[Nieuwejaar]

finally failed after 563 iterations.

[Nieuwejaar]

There are two issues here.

First, the test wasn't being set up correctly. The port wasn't being marked as "nat only", so the packet was triggering an ICMP_UNREACHABLE rather than being dropped. This is the packet seen in the test flake.

Second, there is an issue with the test harness. As written, a test will run until we hit the timeout or until we receive all the packets that test expects to receive. In this case, the error manifests as an ICMP packet we aren't expecting. So, in most cases the test will complete "successfully" as soon as it sees the test packet enter the switch, and will not notice the unexpected ICMP packet exiting the switch a fraction of a second later.

The same timing issue is responsible for the test flake reported in #107.