RFR: 8231260: (dc) DatagramChannel::disconnect changes the port of the local address to 0 (lnx)

[Florian Weimer]

* Daniel Fuchs:

> Hi,
>
> Please find below a fix for:
>
> 8231260: (dc) DatagramChannel::disconnect changes the port of the
>          local address to 0 (lnx)
> https://bugs.openjdk.java.net/browse/JDK-8231260

I remember looking at this a couple of years ago.  I was mostly
concerned about the legacy datagram socket back then, where the
rebinding can fail as well.  Not sure if a bug was field about the
behavioral different between the two socket implementations at the time.

I'm a bit worried about the first hunk:

@@ -873,10 +873,15 @@
                 synchronized (stateLock) {
                     ensureOpen();
                     if (state == ST_CONNECTED)
                         throw new AlreadyConnectedException();
 
+                    // ensure that the socket is bound
+                    if (localAddress == null) {
+                        bindInternal(null);
+                    }
+
                     int n = Net.connect(family,
                                         fd,
                                         isa.getAddress(),
                                         isa.getPort());
                     if (n <= 0)


Doesn't it introduce a race condition where the socket has a port and
queue packets, but the kernel will not perform source address checking
because it is unconnected?

The legacy datagram socket implementation also has logic to deal with
packets that arrive after a disconnect (CVE-2014-6512, the
explicitFilter stuff).  I think DatagramChannelImpl would need something
comparable if we want to paper over the kernel behavior.

Some background:

  <https://www.openwall.com/lists/oss-security/2014/10/17/8>
  <https://www.openwall.com/lists/oss-security/2017/11/06/9>

Thanks,
Florian