CYSDL 3: Solution

This is the solution to the Can You Spot the DeadLock 3.

The problem

No deadlock today, this code does not produce the expected result (27000) with certainty. So it raises an explicit exception, most of the time.

why does it occur?

As you have probably guessed, the general intent is to have three agents, each of them responsible for applying a factor (2,3,5) in a sequential way. Sequence is orchestrated thanks to a step counter and the use of modulo so that each agent acts on the appropriate steps.

Expected
Expected Result

The major failure here is the use of if(!condition) instead of while(!condition).
Improper use of the monitor pattern: agents assume that if they are woken up, it means their condition is fulfilled, which is clearly not the case: they are waken for each forward steps. Afterward, it’s a race condition between threads to get the lock.

Sample result
Sample result

As you can see, neither order nor execution count is secured. This code fails in many ways.

how to fix?

You just need to replace the ‘if(!condition)’ with ‘while(!condition)’ and bang, you’re done.

Actually, writing this code was harder than anticipated. My original plan was to use Pulse instead of PulseAll as I feared that PulseAll was giving too much away. But it turns out that all my attempts led to non working code or too obvious problems.

Feel free to submit your own if you want.

And see you for the next exercise.

Advertisements

Can You Spot the Deadlock 3?

Hello dear readers and coders

Welcome to the third ‘Can You Spot the Deadlock?’ trivia. Today’s exercise is still dedicated to the Monitor Pattern. Next one will explore other horizons.

As usual the code is in C#, it does compile but it will not run as is.

So:

  • What is the issue here?
  • Where does it occur?
  • Can you fix it ? if yes how, if not, why

And please share your ideas through the comment!

using System;
using System.Threading;

namespace SpotTheDeadLock
{
  internal class SpotTheDeadlock3
  {
    readonly object _synchro = new object(); // lock object

    public int _x = 1; // resulting data

    bool _mustStop;
    int _iteration = 0;
    int []_factor={2,3,5};

    // first running thread

    public void Runner1(int stepId) {
      while (true) {
        lock (_synchro) {
          if (_iteration % 3 != stepId && !_mustStop)
            Monitor.Wait(_synchro);

          if (_mustStop)
            return;

          _x = _x * _factor[stepId];
          NextStep();

        } // _synchro
      } // while
    }

  private void NextStep() {
    if (_iteration == 8)
      _mustStop = true;
    else
      _iteration++;
    Monitor.PulseAll(_synchro);
  }

  public static void Runner() {

    SpotTheDeadlock3 agent = new SpotTheDeadlock3();
    Thread thread1 = new Thread(() => agent.Runner1(0));
    Thread thread2 = new Thread(() => agent.Runner1(1));
    Thread thread3 = new Thread(() => agent.Runner1(2));

    thread1.Start();
    thread2.Start();
    thread3.Start();

    thread1.Join();
    thread2.Join();
    thread3.Join();

    if (agent._x != 2 * 2 * 2 * 3 * 3 * 3 * 5 * 5 * 5)
      throw new ApplicationException("Invalid result");
    }

  } // class
} // namespace

Quizz time is closed, solution is available here

CYSDL 2: Solution

This is the solution to the Can You Spot the DeadLock 2.

The problem

This code can deadlock. It actually deadlocks pretty fast on my testbed. The deadlock occurs between the notification mechanism and the unsubscription method in the client class!

Where does it occur?

The notification thread is locked trying to gets its notification through

public class Client: IDisposable {
...
private void OnEvent(object sender, EventArgs arguments) {
lock (this) // notification thread is locked here

The main thread is locked trying to unsubscribe properly
public class EventSource {
...
public event EventHandler EventOccured{
remove {
lock (this) // main thread is locked here
}

The source of the issue is that we have two conflicting resource acquisition paths:
1) The notification path, where first (the lock for) the event is acquired (to prevent modification on the fly of the subscribers list) and then the client (lock) is acquired (to ensure exclusive access to its internals).
2) The unsubscription path wich first acquire the client (to ensure exclusive access to its internals) and then acquire the event for safe unsubscription.

How can it be fixed?

That is the impossible part. There is no proper solution to this problem! That one of the reason I like it so much. But the keyword here is proper. You can resolve the deadlock assuming you are ok to relax the contract a bit.

What is the contract here?


This code (unsuccessful) implements a strong contract: The Client will not receive any notification before the end of the subscription method and it will not receive any notifications after the end of unsubscription method. The first assumption is in fact trivial: how could a client receive notifications from a source it has not subscribed to yet :-). The second one looks similar and quite harmless, it even looks like very desirable. But its a trap: every implementation of this contract offers deadlock opportunities!

You need to relax it a bit and remove the second requirement: you must accept to (potentialy) receive notifications even after a succesful unsubscription. Then you can drop the locks in EventSource.

Can You Spot the Deadlock 2?

Hello dear readers and coders

Welcome to the second ‘Can You Spot the Deadlock?’ trivia. I surely hope you had fun with the first one. Yeah, I know, it was kinda easy. So today, I raise the bar a bit and bring you one of my favorites: the transactional subscription/un subscription pattern.

The code compiles but it will not run, as I did not show the scaffolding code. Suffice to say that there is a notification mechanism that lives in its own thread(s) and some client code logic that rely on the main(starting) thread.

So:

  • What is the issue here?
  • Where does it occur?
  • Can you fix it ? if yes how, if now, why

Problem is now closed, please find the solution

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;

namespace SpotTheDeadLock2
{
///
/// Implements aynchronous notification
///
public class EventSource
{
    private EventHandler _eventHandler;
    ///
    /// notification event
    ///

    public event EventHandler EventOccured
    {
        add {
            // lock to secure subscription
            lock (this){
                _eventHandler += value;
            }
        }
        remove {
            // lock to secure unsubscription
            lock (this) {
                _eventHandler -= value;
            }
        }
    }
    // implementation of notification
    private void TriggerEvent() {
        lock (this) {
            if (_eventHandler != null) {
                _eventHandler(this, new EventArgs());
            }
        }
    }
}

///
/// client implementation
///
public class Client: IDisposable
{
    private EventSource _source;
    ///
    /// Simple constructor
    ///
    ///data source to subsribe to
    public Client(EventSource source) {
        _source = source;
        _source.EventOccured += OnEvent;
    }

    ///
    /// Method in charge of processing events
    ///
    private void OnEvent(object sender, EventArgs arguments) {
        // use lock as we do not have control of calling thread
       lock (this) {
           // do my job
           ...
       }
    }
    // unsubscribe on clean up
    public void Dispose() {
    // use lock to ensure no processing is in progress
        lock (this) {
            _source.EventOccured -= OnEvent;
        }
    }
}
}

CYSDL 1: Solution

As promised, here is the answer to my question
Here the deadlock manifests itself as the main thread stuck on one of the Thread.Join calls and the associated thread’s (_firstRunner or _secondRunner) Monitor.Wait(_synchro) call.

The problem lies with the Monitor.Pulse(_synchro) call, because it wakes only one thread waiting allowing it to end properly. The other thread is still waiting for a signal that will never happen. And ultimately, the main thread is waiting for both threads to end, and we already know that it will not happen.

How to fix: you need to replace Monitor.Pulse(_synchro) by Monitor.PulseAll(_synchro).

This example reminds us that deadlocks do not systematicaly occurs on lock statement and that one must be careful regarding thread termination. Plus, here the problem does not relate to the classical ordering of resource acquisition.

The only random aspect was which of the two runner threads would be locked.

It also raises concern about when and why someone has to worry about using Pulse or PulseAll.

How can it be improved: there is a call to Thread.Sleep that I did place just to make sure that both runner thread where waiting for the signal. This is never the proper way to make a rendez-vous point between threads. But this is neither a trivial matter and adequate rendez-vous code would add a significant chunck of code.

Rendez-vous will probably a topic for another exercise.

How did you fare in this exercise

PS: This an automated post