Be careful using anonymous delegates across threads

Anonymous delegates are incredibly useful, especially in that they can “capture” variables from a parent scope. Used within a single thread, they are very easy to understand. However, if you’re using them to pass data across threads, you need to understand how the variable capture works.

UPDATE: It has been pointed out to me that the premise of this article is wrong. The compiler specifically avoids this problem by hoisting the captured variables into an inner class of which the generated method is also an instance member, and creating a new instance of this class for each call. I’m not sure what bug I ran into that causes the behaviour I denounced, or whether it was simply sloppy handling of reference types on my part – but I feel particularly stupid having vaguely remembered “inner classes” from reading Raymond Chen’s fine explanation of this over a year ago, yet not having checked it before posting this article. Nonetheless, I shall let the article remain as a reminder of the code that this wonderful feature saves you from having to write. Also, the queue/lock-based code may still be useful when one needs data passed to the GTK thread in a guaranteed order.

Consider the following code snippet, in which the “data” variable is captured from the delegate’s parent scope.

void ProcessIncomingData (string data)
{
    Gtk.Application.Invoke (delegate (object sender, EventArgs e) {
            textBuffer.AddText (data);
        });
}

What’s wrong with this?

Let’s expand the anonymous method into something resembling the compiler-generated code.

void ProcessIncomingData (string data)
{
    compilerHoistedVariable = name;
    Gtk.Application.Invoke (new EventHandler (compilerGeneratedMethod));
}

string compilerHoistedVariable = null;

void compilerGeneratedMethod (object sender, EventArgs e)
{
    textBuffer.AddText (compilerHoistedVariable);
}

Notice that the variable has been “hoisted” into the class scope, so that the generated method used for the anonymous delegate can access it. (The actual compiler-generated code is a bit different, encapsulating the hoisted variables in an inner class, but the principle is the same).

Now consider what happens if ProcessIncomingData gets called by the data-processing thread several times before the GTK+ loop gets run and can handle the invocations. Only the most recent value remains in the compiler-hoisted variable, and the earlier data strings are lost.

This isn’t a problem if you’re passing transient state, for example the visibility state of a button. But if you’re passing data, think carefully about the internal mechanics. You may have to implement locking on the shared data structure.

void ProcessIncomingData (string data)
{
    lock (textBufferQueue) {
        textBufferQueue.Enqueue (data);
        if (!queueHandlerRunning) {
            queueHandlerRunning = true;
            Gtk.Application.Invoke (
                delegate (object sender, EventArgs e) {
                    lock (textBufferQueue) {
                        while (textBufferQueue.Count > 0) {
                            textBuffer.AddText (textBufferQueue.Dequeue ());
                        }
                        queueHandlerRunning = false;
                    }
                });
        }
    }
}
bool queueHandlerRunning = false;
Queue<string> textBufferQueue = new Queue<string> ();

This could hardly be described as “pretty”, and the anonymous delegate has grown so big that we may as well move its definition out into a conventional method.

Things are easier if the delegate consumer offers an alternate call allowing you to pass arguments along with the delegate, such as Gtk.Application.Invoke and its Invoke (object sender, System.EventArgs args, System.EventHandler d) overload.

void ProcessIncomingData (string data)
{
    Gtk.Application.Invoke (
        this,
        new dataEventArgs (data),
        delegate (object sender, dataEventArgs e) {
            textBuffer.AddText (dataEventArgs.Data);
        });
}

class dataEventArgs : EventArgs
{
    public string Data;
    public dataEventArgs (string data)
    {
        this.Data = data;
    }
}

This is much more verbose than the method with which we started, but at least you won’t lose any data and don’t have to worry about locking.

Sadly, the anonymous types in C# 3.0 can’t be used to simplify this. For a start, they don’t allow inheritance. One might think that the “object sender” argument could be abused to pass an anonymous type, however in order to leave the scope of a method, the anonymous type must be cast to “object”. The only way to access its properties would be via reflection, which would not only be verbose but would also carry a performance penalty.