connections closed during eBPF initialisation are never removed

[rade]

In a similar vein to #2689...

Apply

diff --git a/probe/endpoint/connection_tracker.go b/probe/endpoint/connection_tracker.go
index 7d6f7ea2..9cd90bcc 100644
--- a/probe/endpoint/connection_tracker.go
+++ b/probe/endpoint/connection_tracker.go
@@ -2,6 +2,7 @@ package endpoint
 
 import (
        "strconv"
+       "time"
 
        log "github.com/Sirupsen/logrus"
        "github.com/weaveworks/scope/probe/endpoint/procspy"
@@ -170,6 +171,8 @@ func (t *connectionTracker) getInitialState() {
                }
        })
 
+       time.Sleep(10 * time.Second)
+
        t.ebpfTracker.feedInitialConnections(conns, seenTuples, processesWaitingInAccept, report.MakeHostNodeID(t.conf.HostID))
 }

and then

1. run nc www.weave.works 80 in one window
2. run scope launch in another window
3. within 10 seconds, in the first window type GET /<RET><RET>, which should terminate the connection.

Now bring up the process topology. This will show a nc process connected to The Internet. Eventually the process will become bare except for the pid:

That node and edge will never go away.

[rade]

The problem here is that we ignore 'close' events until after we've fed all existing connections to the ebpf tracker. So any connections that get closed after we've obtained the initial connection list but before we've fed them to the ebpf tracker, are stashed away by the tracker as open connections and never removed, since the only thing removing them is a 'close' event, which we've missed.