What kind of 'volatile' operation is needed in Double checked locking in .NET

Question

I took the code of DCL from Joe Duffy's book 'Concurrent programming on windows'

class LazyInit<T> where T : class
{
private volatile T m_value;
private object m_sync = new object();
private Func<T> m_factory;
public  LazyInit(Func<T>  factory)  {  m_factory  =  factory;  }
public T value
{
  get
  {
    if (m_value == null)
    {
      lock (m_sync)
      {
        if (m_value == null)
        {
          m_value = m_factory();
        }
      }
    }
    return m_value;
  }
}
}

it is said marking m_value volatile can prevent writes reordering that will leads to other threads getting 'non null object with uninitialized fields'. If the problem happens just because the possible writes reordering, can I just use 'Volatile Write' instead of marking the filed volatile, like below? (This code looks a little awkward for demonstration, I just want to make sure if we can only use volatile write instead)

class LazyInit<T> where T : class
{
private object m_value;
private object m_sync = new object();
private Func<T> m_factory;
public  LazyInit(Func<T>  factory)  {  m_factory  =  factory;  }
public T value
{
  get
  {
    if (m_value == null)
    {
      lock (m_sync)
      {
        if (m_value == null)
        {
          Thread.VolatileWrite(ref m_value, m_factory());
        }
      }
    }
    return (T)m_value;
  }
}
}

A related question is the Interlocked version from the book

class LazylnitRelaxedRef<T> where T : class
{
private volatile T m_value;
private Func<T> m_factory;
public LazylnitRelaxedRef(Func<T> factory) { m_factory = factory; }
public T Value
{
  get
  {
    if (m_value == null)
      Interlocked.CompareExchange(ref  m_value, m_factory(), null);
    return m_value;
  }
}
}

Since the ECMA-CLI specs the 'Interlocked operation perform implicit acquire/release operations', do we still need volatile in this case?

Answer 1

First, messing with volatile is really hard, so don't get too loose with it! But, here is a really close answer to your question, and here is an article that I think everyone should read before using the keyword volatile, and definitely before starting to use VolatileRead, VolatileWrite and MemoryBarrier.

The answer in the first link is: no you don't need to use volatile, you just need to use System.Threading.Thread.MemoryBarrier() RIGHT BEFORE you assign the new value. This is because the release_fence implied when using the volatile keyword makes sure that it gets finished writing out to the main memory, and that no read/write operations can be performed until it's finished.

So, what does Thread.VolatileWrite() do, and does it perform the same functions that we get from the 'volatile' keyword? Well, here's the full code from this function:

public static void VolatileWrite (ref int address, int value)
{
  MemoryBarrier(); address = value;
}

Yes, it calls MemoryBarrier right before it assigns your value, which is sufficient!