Question

I have a Windows Forms app that itself launches different threads to do different kinds of work. Occasionally, ALL threads (including the UI thread) become frozen, and my app becomes unresponsive. I’ve decided it may be a Garbage Collector-related issue, as the GC will freeze all managed threads temporarily. To verify that just managed threads are frozen, I spin up an unmanaged one that writes to a “heartbeat” file with a timestamp every second, and it is not affected (i.e. it still runs):

public delegate void ThreadProc();

[DllImport("UnmanagedTest.dll", EntryPoint = "MyUnmanagedFunction")]
public static extern void MyUnmanagedFunction();

[DllImport("kernel32")]
public static extern IntPtr CreateThread(
    IntPtr lpThreadAttributes,
    uint dwStackSize,
    IntPtr lpStartAddress,
    IntPtr lpParameter,
    uint dwCreationFlags,
    out uint dwThreadId);    

uint threadId;
ThreadProc proc = new ThreadProc(MyUnmanagedFunction);
IntPtr functionPointer = Marshal.GetFunctionPointerForDelegate(proc);
IntPtr threadHandle = CreateThread(IntPtr.Zero, 0, functionPointer, IntPtr.Zero, 0, out threadId);

My Question is: how can I simulate this situation, where all managed threads are suspended but unmanaged ones keep on spinning?

My first stab:

private void button1_Click(object sender, EventArgs e) {
    Thread t = new Thread(new ThreadStart(delegate {
        new Hanger();
        GC.Collect(2, GCCollectionMode.Forced);
    }));
    t.Start();
}
class Hanger{
    private int[] m_Integers = new int[10000000];
    public Hanger() { }
    ~Hanger() { Console.WriteLine("About to hang...");

    //This doesn't reproduce the desired behavior
    //while (true) ;

    //Neither does this
    //Thread.Sleep(System.Threading.Timeout.Infinite); 
    }
}

Thanks in advance!!

Answer 1

The issue DID in fact stem from the Garbage Collector. After many days of debugging and analyzing memory dumps with WinDbg, we realized that there was a deadlock situation, but induced by the GC collecting concurrently. Changing the GC to collect non-concurrently fixed our problem.