The PVS-Studio analyzer can detect incorrect variants of the "double-checked locking" pattern implementation in C# programs. Sometimes users are not aware of the danger in the code indicated by the analyzer and the ways how to correct it. Therefore, we will illustrate with a practical example what the error detected by the V3054 warning can look like and how to fix the code.
The PVS-Studio analyzer can detect the error of unsafe implementation of the double checked locking pattern. Double-checked locking is a pattern meant for reducing the overhead of locking. First, the locking condition is checked without synchronization. And only if the condition is met, the thread will try to get a lock. Thus, the locking would be executed only if it was really necessary.
Code that implements this pattern may be written not neatly enough. It can be especially upsetting that such code might fail very rarely, which makes it difficult to identify the problem in code. So even if it seems to you that the program is working as intended and the code is written correctly, you should carefully pay attention to the corresponding analyzer warning.
If suspicious code is detected, PVS-Studio will issue the warning: V3054 [CWE-609] Potentially unsafe double-checked locking. Use volatile variable(s) or synchronization primitives to avoid this.
Let's look at a real example of what a similar error looks like. The following code fragment is taken from the RunUO project. Recently, we've written about the check of this project in this article.
private Packet m_RemovePacket;
....
private object _rpl = new object();
public Packet RemovePacket
{
get
{
if (m_RemovePacket == null)
{
lock (_rpl)
{
if (m_RemovePacket == null)
{
m_RemovePacket = new RemoveItem(this);
m_RemovePacket.SetStatic();
}
}
}
return m_RemovePacket;
}
}The PVS-Studio analyzer issues the warning: V3054 Potentially unsafe double-checked locking. Use volatile variable(s) or synchronization primitives to avoid this. Item.cs 1624
As can be seen from the above code, double checked locking was applied to implement the singleton pattern. When attempting to get the Packet class instance and addressing the RemovePacket property, the getter checks the m_RemovePacket field for null. If the check is successful, we get into the body of the lock operator, where the field m_RemovePacket gets initialized. The plot thickens when the main thread has already initialized the m_RemovePacket variable through the constructor, but hasn't called the SetStatic() method yet. In theory, another thread can access the RemovePacket property at this very awkward moment. The check of m_RemovePacket for null will fail and the caller thread will get the reference to a half ready-to-use object. To solve this problem, we can create an intermediate variable of Packet class in the body of the lock operator, initialize the variable via the constructor and the SetStatic() method, and after assign it to the m_RemovePacket variable. This way, the body of the lock operator might look as follows:
lock (_rpl)
{
if (m_RemovePacket == null)
{
Packet instance = new RemoveItem(this);
instance.SetStatic();
m_RemovePacket = instance;
}
}It seems that the problem has been fixed and the code will work as expected. But not so fast.
Here's another thing: the analyzer offers to use the volatile keyword for a reason. In the release version of the program, the compiler might optimize and reorder calling lines of the SetStatic() method and assignment of the instance variable to the m_RemovePacket field (from the compiler's point of view, program semantics won't break). Here we get back to the point where we started - the m_RemovePacket variable might be uninitialized. We can't say exactly when this reordering may occur. We are even not sure if it happens at all, as the CLR version, the architecture of the used processor and other factors might affect it. It's still worth preventing this scenario. In this regard, one of the solutions (not the most productive) will be usage of the keyword volatile. The variable declared with the volatile modifier won't be object to displacements during compiler optimizations. The final code version might look as follows:
private volatile Packet m_RemovePacket;
....
private object _rpl = new object();
public Packet RemovePacket
{
get
{
if (m_RemovePacket == null)
{
lock (_rpl)
{
if (m_RemovePacket == null)
{
Packet instance = new RemoveItem(this);
instance.SetStatic();
m_RemovePacket = instance;
}
}
}
return m_RemovePacket;
}
}In some cases, it's undesirable to use a volatile field due to some cost of accessing this field. Let's not dwell on this issue, noting simply that in this example, the atomic field writing is needed only once (when first accessing the property). However, volatile field declaration will lead to the fact that the compiler will atomically perform its each reading and writing, which might be non-optimal in terms of performance.
Therefore, let's consider another way to avoid this analyzer warning. We can use the Lazy<T> type for the m_RemovePacket backing field instead of double checked locking. As a result, we'll get rid of potential cost of declaring a volatile field. In this case, the body of the getter can be replaced by the initializing method, which will be passed to the constructor of the Lazy<T> instance:
private Lazy<Packet> m_RemovePacket = new Lazy<Packet>(() =>
{
Packet instance = new RemoveItem(this);
instance.SetStatic();
return instance;
}, LazyThreadSafetyMode.ExecutionAndPublication);
....
public Packet RemovePacket
{
get
{
return m_RemovePacket.Value;
}
}The initializing method will be called only once when first accessing the instance of the Lazy type. In doing so, the Lazy<T> type will ensure thread security in case of simultaneous multi-thread access to a property. The thread security mode is controlled by the second parameter of the Lazy constructor.