As you may be aware, if a host receives a packet for a UDP port that is not currently bound, it may send back an ICMP “Port Unreachable” message. Whether or not it does this is dependent on the firewall, private/public settings, etc. On localhost, however, it will pretty much always send this packet back.
In your server code, you are calling SendAsync to old clients, which prompts these “port unreachable” messages.
Now, on Windows (and only on Windows), by default, a received ICMP Port Unreachable message will close the UDP socket that sent it; hence, the next time you try to receive on the socket, it will throw an exception because the socket has been closed by the OS.
Obviously, this causes a headache in the multi-client, single-server socket set-up you have here, but luckily there is a fix:
You need to utilise the not-often-required SIO_UDP_CONNRESET Winsock control code, which turns off this built-in behaviour of automatically closing the socket.
I don’t believe this ioctl code is available in the dotnet IoControlCodes type, but you can define it yourself. If you put the following code at the top of your server repro, the error no longer gets raised.
const uint IOC_IN = 0x80000000U;
const uint IOC_VENDOR = 0x18000000U;
/// <summary>
/// Controls whether UDP PORT_UNREACHABLE messages are reported.
/// </summary>
const int SIO_UDP_CONNRESET = unchecked((int)(IOC_IN | IOC_VENDOR | 12));
var udpClient = new UdpClient(9001);
udpClient.Client.IOControl(SIO_UDP_CONNRESET, new byte[] { 0x00 }, null);
Note that this ioctl code is only supported on Windows (XP and later), not on Linux, since it is provided by the Winsock extensions. Of course, since the described behavior is only the default behavior on Windows, this omission is not a major loss. If you are attempting to create a cross-platform library, you should cordon this off as Windows-specific code.