Re: Finalizer deadlock

[Bryce McKinlay <mckinlay at redhat dot com>]

Gladish, Jacob wrote:

> This was proposed about two months ago:
>
>  
>
> > Yes, we could rewrite it using _Jv_CondWait/_Jv_CondNotify and other
> > thread primitives defined in the thread headers.
> >
> > I'm also not sure this peice of code in FinalizerThread is safe:
> > public static void finalizerReady ()
> > {
> >  synchronized (lock)
> >    {
> >  if (! thread_started)
> >    runFinalizers ();
> >
> > Isn't there a race here? Unlikely to happen, perhaps, but it seems like
> >    
> if >a finalizer needs to be run before the finalizer thread has finished
>  
>
> > starting, we'll try to run finalizers from the inside the GC's thread.
> >
> > Regards
> >
> > Bryce
> >
> >    
>
>
> Is this that case? The two threads are deadlocked. The finalizer is
> waiting for a monitor on the PlainSocketImpl. 

Hmm, looking at this more, I think a race condition is still possible  - 
although very unlikely ...

// Main loop for the finalizer thread.
void
gnu::gcj::runtime::FinalizerThread::run ()
{
 while (true)
   {
     _Jv_MutexLock (&mutex);
     if (! finalizer_ready)
   _Jv_CondWait (&condition, &mutex, 0, 0);
     finalizer_ready = false;
     _Jv_MutexUnlock (&mutex);
     _Jv_RunFinalizers ();
   }
}

If the finalizer thread were suspended by a GC cycle being invoked from 
another thread, while it holds the mutex, then a deadlock could occur if 
that GC then invokes finalizerReady().

Perhaps we need some kind of "GC critical section" call to ensure that 
the GC does not attempt to suspend the finalizer thread while the lock 
is held. Or, perhaps there is an easier solution - could this be 
implemented purely with atomic operations? Java finalization sure is 
tricky to implement (reliably)...

Still, I think this is very unlikely to actually happen (note that the 
mutex is released by _Jv_CondWait), and is unlikely to be the cause of 
the bug you saw.

Regards

Bryce