If a master fails to ping a slave, it should be hard-disconnected.

changelog.html

@@ -43,6 +43,8 @@ Upcoming changes</a>
   <li class=bug>
     Fixed a bug in crontab "day of week" handling in locales where a week starts from Monday.
     (<a href="http://issues.hudson-ci.org/browse/HUDSON-8401">issue 8401</a>)
+  <li class=bug>
+    If a master fails to ping a slave, it should be hard-disconnected.
 </ul>
 </div><!--=TRUNK-END=-->
 

core/src/main/java/hudson/TcpSlaveAgentListener.java

@@ -116,6 +116,12 @@ public final class TcpSlaveAgentListener extends Thread {
             // the loop eventually terminates when the socket is closed.
             while (true) {
                 Socket s = serverSocket.accept();
+
+                // this prevents a connection from silently terminated by the router in between or the other peer
+                // and that goes without unnoticed. However, the time out is often very long (for example 2 hours
+                // by default in Linux) that this alone is enough to prevent that.
+                s.setKeepAlive(true);
+
                 new ConnectionHandler(s).start();
             }
         } catch (IOException e) {

core/src/main/java/hudson/node_monitors/ResponseTimeMonitor.java

@@ -25,6 +25,7 @@ package hudson.node_monitors;
 
 import hudson.Util;
 import hudson.Extension;
+import hudson.model.Descriptor.FormException;
 import hudson.slaves.OfflineCause;
 import hudson.model.Computer;
 import hudson.remoting.Callable;
@@ -67,8 +68,14 @@ public class ResponseTimeMonitor extends NodeMonitor {
                 d = new Data(old,-1L);
             }
 
-            if(d.hasTooManyTimeouts() && markOffline(c,d))
+            if(d.hasTooManyTimeouts() && !isIgnored()) {
+                // unlike other monitors whose failure still allow us to communicate with the slave,
+                // the failure in this monitor indicates that we are just unable to make any requests
+                // to this slave. So we should severe the connection, as opposed to marking it temporarily
+                // off line, which still keeps the underlying channel open.
+                c.disconnect(d);
                 LOGGER.warning(Messages.ResponseTimeMonitor_MarkedOffline(c.getName()));
+            }
             return d;
         }
 

core/src/main/resources/hudson/node_monitors/Messages.properties

@@ -25,7 +25,7 @@ ClockMonitor.DisplayName=Clock Difference
 DiskSpaceMonitor.MarkedOffline=Making {0} offline temporarily due to the lack of disk space
 DiskSpaceMonitor.DisplayName=Free Disk Space
 ResponseTimeMonitor.DisplayName=Response Time
-ResponseTimeMonitor.MarkedOffline=Making {0} offline temporarily because it''s not responding
+ResponseTimeMonitor.MarkedOffline=Making {0} offline because it''s not responding
 ResponseTimeMonitor.TimeOut=Time out for last {0} try
 SwapSpaceMonitor.DisplayName=Free Swap Space
 TemporarySpaceMonitor.DisplayName=Free Temp Space
\ No newline at end of file