ServeBackground returns before UDP connection is added to sipgo's pool, causing "bind: address already in use" on Register

[jmarble]

When calling Register() immediately after ServeBackground() returns, the REGISTER transaction fails with:

client transcation failed to request connection: listen udp 127.0.0.1:35256: bind: address already in use

Note: "transcation" in the error string is a typo in sipgo (transaction).

Root cause:

In diago.serve(), the ListenReadyFuncCtxValue callback fires after net.ListenUDP succeeds (server.go:129) but before srv.tp.ServeUDP() runs (server.go:130), which is where pool.Add happens (transport_udp.go:61):

// server.go ListenAndServe
udpConn, err := net.ListenUDP(network, laddr)  // binds port
listenReadyCtx(ctx, network, udpConn.LocalAddr().String())  // fires readyCh here
return srv.tp.ServeUDP(udpConn)  // pool.Add happens inside here

So ServeBackground returns while the connection pool is still empty. When Register() runs, ClientRequestConnection calls GetConnection(laddr) -> returns nil -> calls CreateConnection -> tries ListenPacket on the same port -> OS rejects it.

The race is acknowledged in comments at diago.go:874:

// Checking ports with ListenPorts is racy with ListenAndServe
// In case UDP, we want to have this port reused.
bindPort = tran.BindPort

Reproduction:

dg := diago.NewDiago(ua, diago.WithTransport(diago.Transport{
    Transport:      "udp",
    BindHost:       "127.0.0.1",
    BindPort:       0, // ephemeral
    RewriteContact: true,
}))

ctx, cancel := context.WithCancel(context.Background())
dg.ServeBackground(ctx, handler)

// This can fail with "address already in use"
dg.Register(ctx, registrar, diago.RegisterOptions{...})

Suggested fix:

Move the readyCh signal so it fires after pool.Add rather than after ListenUDP. One approach: have ServeUDP/TransportUDP.Serve accept an optional ready callback that fires after pool.Add:

// transport_udp.go
func (t *TransportUDP) Serve(conn net.PacketConn, handler MessageHandler) error {
    c := &UDPConnection{...}
    t.pool.Add(c.PacketAddr, c)  // add to pool first
    // signal ready here, after pool is populated
    t.readListenerConnection(c, c.PacketAddr, handler)
    return nil
}

Or alternatively, listenReadyCtx could be called from within ServeUDP after pool.Add.

Workaround:

Retry Register() when the error contains "address already in use":

for attempt := 1; attempt <= 3; attempt++ {
    err := dg.Register(ctx, registrar, opts)
    if err == nil {
        break
    }
    if strings.Contains(err.Error(), "address already in use") && attempt < 3 {
        time.Sleep(time.Duration(attempt*50) * time.Millisecond)
        continue
    }
    return err
}

Versions: diago v0.27.0, sipgo v1.2.0

[jmarble]

sorry, technically not a diago bug, but rather a sipgo bug.

[emiago]

Yeah I am already aware of this problem. It is even issue when you try check ports on transport layer after this call.
Will add some priority to solve.

Yes we need to solve on SIPgo or diago should use raw API for Serve each transport.