.Net 4. ThreadLocal<> implements IDisposable. But it seems that calling Dispose() doesn't actually release references to thread local objects being held.
This code reproduces the problem:
using System;
using System.Collections.Generic;
using System.Collections.Concurrent;
using System.Linq;
using System.Threading;
namespace ConsoleApplication2
{
class Program
{
class ThreadLocalData
{
// Allocate object in LOH
public int[] data = new int[10 * 1024 * 1024];
};
static void Main(string[] args)
{
// Stores references to all thread local object that have been created
var threadLocalInstances = new List<ThreadLocalData>();
ThreadLocal<ThreadLocalData> threadLocal = new ThreadLocal<ThreadLocalData>(() =>
{
var ret = new ThreadLocalData();
lock (threadLocalInstances)
threadLocalInstances.Add(ret);
return ret;
});
// Do some multithreaded stuff
int sum = Enumerable.Range(0, 100).AsParallel().Select(
i => threadLocal.Value.data.Sum() + i).Sum();
Console.WriteLine("Sum: {0}", sum);
Console.WriteLine("Thread local instances: {0}", threadLocalInstances.Count);
// Do our best to release ThreadLocal<> object
threadLocal.Dispose();
threadLocal = null;
Console.Write("Press R to release memory blocks manually or another key to proceed: ");
if (char.ToUpper(Console.ReadKey().KeyChar) == 'R')
{
foreach (var i in threadLocalInstances)
i.data = null;
}
// Make sure we don't keep the references to LOH objects
threadLocalInstances = null;
Console.WriteLine();
// Collect the garbage
GC.Collect();
GC.WaitForPendingFinalizers();
GC.Collect();
Console.WriteLine("Garbage collected. Open Task Manager to see memory consumption.");
Console.Write("Press any key to exit.");
Console.ReadKey();
}
}
}
Thread local data stores a reference to a large object. GC doesn't collect these large objects if references are not nulled manually. I used Task Manager to observe memory consumption. I also run memory profiler. I made a snapshot after garbage was collected. The profiler showed that leaked object is rooted by GCHandle and was allocated with here:
mscorlib!System.Threading.ThreadLocal<T>.GenericHolder<U,V,W>.get_Boxed()
mscorlib!System.Threading.ThreadLocal<T>.get_Value()
ConsoleApplication2!ConsoleApplication2.Program.<>c__DisplayClass3.<Main>b__2( int ) Program.cs
That seems to be a flaw in ThreadLocal<> design. The trick with storing all allocated objects for further cleanup is ugly. Any ideas on how to work around that?