On Mon, 29 May 2006 15:21:36 +0100, Matthew Toseland wrote: > Immediate priorities: > > 1a. Do not do any (synchronous) DNS lookups on PacketSender. 1b. If: > - We are on a Sun/Blackdown JVM (prior to 1.6??) - We are on Linux > - LD_ASSUME_KERNEL isn't set > Then we are vulnerable to the EvilJVMBug, and we must firstly, institute a > critical priority UserAlert explaining the situation, and secondly, refuse > all requests and refuse to send any requestsd.
Before doing anything so drastic, there is one more thing you might want to try. I examined PacketSender and noticed that it constructs the sender thread inside the object constructor, giving it a pointer to the object under construction, and then starts it later in the start() method. Now, suppose that some thread-related resources are allocated at the construct phase and not in the start phase ? That might lead to some confusion inside the JVM. I'm suggesting this because in the past I've observed very strange behavior on programs that did any violations of Java Memory Model (like NullPointerExceptions being thrown just because I switched to an incremental garbage collector). So, try creating the thread object inside the start() method, and not in the object constructor. It *shouldn't* make any difference, but it's worth the try before trying any more drastic solutions. Also, did you have any native libraries loaded when the bug occurred - and if yes, could it possibly be caused by some kind of buffer overrun there ? And yes, Java is quite buggy - I especially love how the documentation states that Graphics.drawImage() method "returns immediately in all cases," when in reality that "immediately" may be up to several seconds long (not a garbage collection pause, since the other threads keep on running) :(. Maybe you should consider reimplementing Freenet in pure ANSI C ?-)
