I'm experimenting with threading in C#, and I've created the following class as a result. I've tried to avoid any cases of race conditions, yet a deadlock occurs on use.
The class uses two different locks, one spinlock for straightforward operations, and additionally a Monitor lock to wait in case no object is ready. I originally used EventWaitHandle, but discovered that race conditions were inevitable due to WaitOne/Set precedence.
Note that Monitor.Pulse could not precede Monitor.Wait, so what else could cause a deadlock? In the case where 5 threads use a TestPool class with a capacity of 4, the deadlock always occurs at SpinLock at an irregular moment.
internal class TestPool<T> where T : class
{
private int capacity;
private int unitPos;
private int waitUnitPos;
private int waitCount;
private int lockState;
private object lockObj;
private T[] units;
private Func<T> unitFactory;
public TestPool(int capacity, Func<T> unitFactory)
{
this.lockObj = new object();
this.unitFactory = unitFactory;
Init(capacity);
}
public T Fetch()
{
T unit;
Lock();
unit = (unitPos != capacity) ? units[unitPos++] : Wait();
Unlock();
return unit;
}
public void Store(T unit)
{
Lock();
if (waitCount == 0)
{
units[--unitPos] = unit;
}
else
{
Pulse(unit);
}
Unlock();
}
private T Wait()
{
waitCount++;
lock (lockObj)
{
Unlock();
Monitor.Wait(lockObj);
Lock();
return units[--waitUnitPos];
}
}
private void Pulse(T unit)
{
waitCount--;
units[waitUnitPos++] = unit;
lock (lockObj)
{
Monitor.Pulse(lockObj);
}
}
private void Lock()
{
if (Interlocked.CompareExchange(ref lockState, 1, 0) != 0)
{
SpinLock();
}
}
private void SpinLock()
{
SpinWait spinWait = new SpinWait();
do
{
spinWait.SpinOnce();
}
while (Interlocked.CompareExchange(ref lockState, 1, 0) != 0);
}
private void Unlock()
{
Interlocked.Exchange(ref lockState, 0);
}
private void Init(int capacity)
{
T[] tx = new T[capacity];
for (int i = 0; i < capacity; i++)
{
tx[i] = unitFactory.Invoke();
}
units = tx;
this.capacity = capacity;
}
}
Fixed it. I had to place the following code outside the Monitor lock.
Lock();
return units[--waitUnitPos];
Related
I am working on my own multithreading for my algorithm independed pathfinding for unity. However, when I am executing two the same class I get a memory leak and when only executing one instance I am having no issues. I really want to use at least two threads if it is necessary.
Below is the class I have issues with. Keep in mind, that two independend threads will have to execute parts of this script. AddJob can be called from the main unity thread but will most likely be called from another update thread for the agents.
namespace Plugins.PathFinding.Threading
{
internal class PathFindingThread
{
private Thread m_Worker;
private volatile Queue<CompletedProcessingCallback> m_CallbackQueue;
private volatile Queue<IAlgorithm> m_QueuedTasks;
internal int GetTaskCount
{
get
{
return m_QueuedTasks.Count;
}
}
internal PathFindingThread()
{
m_Worker = new Thread(Run);
m_CallbackQueue = new Queue<CompletedProcessingCallback>();
m_QueuedTasks = new Queue<IAlgorithm>();
}
private void Run()
{
Debug.Log("<b><color=green> [ThreadInfo]:</color></b> PathFinding Thread Started ");
try
{
while(true)
{
if (m_QueuedTasks.Count > 0)
{
IAlgorithm RunningTask = m_QueuedTasks.Dequeue();
RunningTask.FindPath(new IAlgorithmCompleted(AddCallback));
}
else
break;
}
Debug.Log("<b><color=red> [ThreadInfo]:</color></b> PathFinding Worker is idle and has been Stopped");
}
catch(Exception)
{
Debug.Log("<b><color=red> [ThreadInfo]:</color></b> PathFinding thread encountred an error and has been aborted");
}
}
internal void AddJob(IAlgorithm AlgorithmToRun)
{
m_QueuedTasks.Enqueue(AlgorithmToRun);
//Debug.Log("Added Job To Queue");
}
private void AddCallback(CompletedProcessingCallback callback)
{
m_CallbackQueue.Enqueue(callback);
}
private void Update()
{
if (m_CallbackQueue.Count > 0)
{
if (m_CallbackQueue.Peek().m_Callback != null) { }
m_CallbackQueue.Peek().m_Callback.Invoke(m_CallbackQueue.Peek().m_Path);
m_CallbackQueue.Dequeue();
}
if (m_Worker.ThreadState != ThreadState.Running && m_QueuedTasks.Count != 0)
{
m_Worker = new Thread(Run);
m_Worker.Start();
}
}
}
internal delegate void IAlgorithmCompleted(CompletedProcessingCallback callback);
internal struct CompletedProcessingCallback
{
internal volatile FindPathCompleteCallback m_Callback;
internal volatile List<GridNode> m_Path;
}
}
namespace Plugins.PathFinding
{
internal enum TypeOfNode
{
Ground,
Air
}
//used to store location information since array can only take rounded numbers
internal struct Position
{
internal int x;
internal int y;
internal int z;
}
internal class GridNode
{
internal Position M_PostitionInGrid { get; private set; }
internal Vector3 M_PostitionInWorld { get; private set; }
internal TypeOfNode M_type { get; private set; }
internal bool m_IsWalkable = true;
internal GridNode m_ParrentNode;
internal int Hcost;
internal int Gcost;
internal int Fcost { get { return Hcost + Gcost; } }
internal GridNode(Position postion , Vector3 WorldPosition)
{
M_PostitionInGrid = postion;
m_IsWalkable = true;
M_PostitionInWorld = WorldPosition;
}
}
}
internal delegate void FindPathCompleteCallback(List<GridNode> Path);
internal abstract class IAlgorithm
{
protected GridNode m_SavedStart;
protected GridNode m_SavedTarget;
protected List<GridNode> m_LocatedPath;
protected FindPathCompleteCallback m_Callback;
internal FindPathCompleteCallback GetCallback
{
get
{
return m_Callback;
}
}
protected PathFindingGrid m_grid;
internal abstract void FindPath(IAlgorithmCompleted callback);
protected abstract List<GridNode> CreatePath(PathFindingGrid Grid, GridNode Start, GridNode Target);
protected abstract List<GridNode> RetracePath(GridNode start, GridNode target);
}
namespace Plugins.PathFinding.Astar
{
internal class AstarFinder : IAlgorithm
{
//construction of the Algorithm
internal AstarFinder(GridNode start, GridNode target, FindPathCompleteCallback Callback)
{
m_SavedStart = start;
m_SavedTarget = target;
m_Callback = Callback;
m_LocatedPath = new List<GridNode>();
m_grid = PathFindingGrid.GetInstance;
}
//function to start finding a path
internal override void FindPath(IAlgorithmCompleted callback)
{
//running Algorithm and getting the path
m_LocatedPath = CreatePath(PathFindingGrid.GetInstance, m_SavedStart, m_SavedTarget);
callback.Invoke(
new CompletedProcessingCallback()
{
m_Callback = m_Callback,
m_Path = m_LocatedPath
});
}
//Algorithm
protected override List<GridNode> CreatePath(PathFindingGrid Grid, GridNode Start, GridNode Target)
{
if(Grid == null ||
Start == null ||
Target == null)
{
UnityEngine.Debug.Log("Missing Parameter, might be outside of grid");
return new List<GridNode>();
}
List<GridNode> Path = new List<GridNode>();
List<GridNode> OpenSet = new List<GridNode>();
List<GridNode> ClosedSet = new List<GridNode>();
OpenSet.Add(Start);
int Retry = 0;
while (OpenSet.Count > 0)
{
if(Retry > 3000 || Grid == null)
{
UnityEngine.Debug.Log("Path Inpossible Exiting");
break;
}
GridNode CurrentNode = OpenSet[0];
for (int i = 0; i < OpenSet.Count; i++)
{
if(OpenSet[i].Fcost < CurrentNode.Fcost || OpenSet[i].Fcost == CurrentNode.Fcost && OpenSet[i].Hcost < CurrentNode.Hcost)
{
CurrentNode = OpenSet[i];
}
}
OpenSet.Remove(CurrentNode);
ClosedSet.Add(CurrentNode);
if(CurrentNode == Target)
{
Path = RetracePath(CurrentNode,Start);
break;
}
GridNode[] neighbour = Grid.GetNeighbouringNodes(CurrentNode);
for (int i = 0; i < neighbour.Length; i++)
{
if (!neighbour[i].m_IsWalkable || ClosedSet.Contains(neighbour[i]))
continue;
int CostToNeighbour = CurrentNode.Gcost + Grid.GetDistance(CurrentNode, neighbour[i]);
if(CostToNeighbour < neighbour[i].Gcost || !OpenSet.Contains(neighbour[i]))
{
neighbour[i].Gcost = CostToNeighbour;
neighbour[i].Hcost = Grid.GetDistance(neighbour[i], Target);
neighbour[i].m_ParrentNode = CurrentNode;
if (!OpenSet.Contains(neighbour[i]))
OpenSet.Add(neighbour[i]);
}
}
Retry++;
}
return Path;
}
//retracing the path out of a node map
protected override List<GridNode> RetracePath(GridNode start, GridNode target)
{
List<GridNode> Output = new List<GridNode>();
GridNode current = start;
while(current != target)
{
Output.Add(current);
current = current.m_ParrentNode;
}
Output.Reverse();
return Output;
}
}
}
This shows the core of your code made thread safe.
internal class PathFindingThread
{
Task m_Worker;
ConcurrentQueue<CompletedProcessingCallback> m_CallbackQueue;
ConcurrentQueue<IAlgorithm> m_QueuedTasks;
internal int GetTaskCount
{
get
{
return m_QueuedTasks.Count;
}
}
internal PathFindingThread()
{
m_CallbackQueue = new ConcurrentQueue<CompletedProcessingCallback>();
m_QueuedTasks = new ConcurrentQueue<IAlgorithm>();
m_Worker = Task.Factory.StartNew(() =>
{
while (true)
{
IAlgorithm head = null;
if (m_QueuedTasks.TryDequeue(out head))
{
head.FindPath(new IAlgorithmCompleted(AddCallback));
}
else
{
Task.Delay(0);
}
}
});
}
internal void AddJob(IAlgorithm AlgorithmToRun)
{
m_QueuedTasks.Enqueue(AlgorithmToRun);
}
private void AddCallback(CompletedProcessingCallback callback)
{
m_CallbackQueue.Enqueue(callback);
}
private void Update()
{
CompletedProcessingCallback cb = null;
if (m_CallbackQueue.TryDequeue(out cb))
{
cb.m_Callback.Invoke(cb.m_Path);
}
}
}
Volatile is only good for changing the value of the field - not calling methods on a collection that is referenced by the field.
You propably do not need to have Volatile in CompletedProcessingCallback, but it depends where else this is used. Certainly having volatile on a struct field is a bad smell.
Resolve these thread issues first, then see if you still have the problem.
I'm experimenting with locks that don't require atomic instructions. Peterson's algorithm seemed like the simplest place to start. However, with enough iterations, something goes wrong.
Code:
public class Program
{
private static volatile int _i = 0;
public static void Main(string[] args)
{
for (int i = 0; i < 1000; i++)
{
RunTest();
}
Console.Read();
}
private static void RunTest()
{
_i = 0;
var lockType = new PetersonLock();
var t1 = new Thread(() => Inc(0, lockType));
var t2 = new Thread(() => Inc(1, lockType));
t1.Start();
t2.Start();
t1.Join();
t2.Join();
Console.WriteLine(_i);
}
private static void Inc(int pid, ILock lockType)
{
try
{
for (int i = 0; i < 1000000; i++)
{
lockType.Request(pid);
_i++;
lockType.Release(pid);
}
}
catch (Exception ex)
{
Console.WriteLine(ex);
}
}
}
public interface ILock
{
void Request(int pid);
void Release(int pid);
}
public class NoLock : ILock
{
public void Request(int pid) { }
public void Release(int pid) { }
}
public class StandardLock : ILock
{
private object _sync = new object();
public void Request(int pid)
{
Monitor.Enter(_sync);
}
public void Release(int pid)
{
Monitor.Exit(_sync);
}
}
public class PetersonLock : ILock
{
private volatile bool[] _wantsCs = new bool[2];
private volatile int _turn;
public void Request(int pid)
{
int j = pid == 1 ? 0 : 1;
_wantsCs[pid] = true;
_turn = j;
while (_wantsCs[j] && _turn == j)
{
Thread.Sleep(0);
}
}
public void Release(int pid)
{
_wantsCs[pid] = false;
}
}
When I run this, I consistently get < 2,000,000. What's going on?
The problem here is these two statements:
_wantsCs[pid] = true;
_turn = j;
The memory model of .NET and C# allows these two writes to be reordered.
To force them to not be reordered, add a memory barrier between them:
_wantsCs[pid] = true;
Thread.MemoryBarrier();
_turn = j;
and you will get the expected output every time.
Note that this very problem is described on the Wikipedia page for Peterson's Algorithm in the note section (shortened down here, go read the article for the full notes):
Notes
...
Most modern CPUs reorder memory accesses to improve execution efficiency (see memory ordering for types of reordering allowed). Such processors invariably give some way to force ordering in a stream of memory accesses, typically through a memory barrier instruction. Implementation of Peterson's and related algorithms on processors which reorder memory accesses generally requires use of such operations to work correctly to keep sequential operations from happening in an incorrect order.
(my emphasis)
I have implemented a simple task to create a fixed sized list that allows concurrent writes and can dump the latest snapshot of items in the list at any time.
Here is my implementation. The offset will increase atomically for each thread and reset if reaches to the size of the list. Different threads should have isolated access to each section of the array.
My question is when I call Dump(), the first few items are not stored in the list. Also, is there a Interlocked function that can do both atomic increase and reset, so I don't have to create a locker object and a lock block? Thanks.
public static void Main(string[] args)
{
ConcurrentCircularFixedList<int> list = new ConcurrentCircularFixedList<int>(20);
Enumerable.Range(1, 30).AsParallel().Select(nu => list.Enqueu(nu)).ToList();
}
public class ConcurrentCircularFixedList<T>
{
private int _size;
private int _offset;
private sealed object _locker = new Object();
privateT[] _list;
public ConcurrentCircularFixedList(int size)
{
_size = size;
_offset = 0;
_list = new T[_size];
}
public int Enqueu(T item)
{
_list[_offset] = item;
lock(_locker)
{
Debug.Write("B " + _offset);
_offset += 1;
if(_offset == _size)
_offset = 0;
Debug.Write("A " + _offset + "\n");
}
return _offset;
}
public T[] Dump()
{
return _list.ToArray();
}
}
Here's a small version of a lock-free list that copies on write. The performance characteristics should be clearly understood before using it. It's expensive when you have many writers or the list is large. Reads are synchronization free since the list is effectively immutable. This could be improved in various ways of course but you get the idea. In effect it sacrifices some memory pressure and slower writes for having zero cost reads.
public class CopyWriteList<T>
{
private volatile List<T> list;
public CopyWriteList()
{
list = new List<T>();
}
public CopyWriteList(int capacity)
{
list = new List<T>(capacity);
}
public T this[int index]
{
get { return list[index]; }
set { Replace(x => x[index] = value); }
}
public void Clear()
{
Replace(x => x.Clear());
}
public void Add(T item)
{
Replace(x => x.Add(item));
}
//Etc....
private void Replace(Action<List<T>> action)
{
List<T> current;
List<T> updated;
do
{
current = list;
updated = new List<T>(current);
action(updated);
} while (Interlocked.CompareExchange(ref list, updated, current) != current);
}
public List<T> GetSnapshot()
{
return list;
}
}
Alternatively here's a fixed version of your code. Note that there is added contention between both readers and writers. Performance could suffer because of it (like the ever expensive context switching).
public class ConcurrentCircularFixedList<T>
{
private readonly int _size;
private int _offset;
private readonly object _locker = new Object();
private readonly T[] _list;
public ConcurrentCircularFixedList(int size)
{
_size = size;
_offset = 0;
_list = new T[_size];
}
public int Enqueue(T item)
{
lock (_locker)
{
_list[_offset] = item;
Debug.Write("B " + _offset);
_offset += 1;
if (_offset == _size)
_offset = 0;
Debug.Write("A " + _offset + "\n");
return _offset;
}
}
public T[] Dump()
{
lock (_locker)
return _list.ToArray();
}
}
I have one thread, that is sending data stored in a buffer of type List< string> via tcp. Another thread is writing into the buffer. As I am not very familiar with c# I'd like to know how I should use lock or Mutex correctly.
This is the code I'd like to use eventually:
while(buffer.isLocked())
{
buffer.wait();
}
buffer.lockBuffer();
buffer.add(tcpPacket);
buffer.unlockBuffer();
buffer.notify();
This is my current code. I hope someone can help me complete it.
public class Buffer
{
private Mutex mutex;
private List<string> buffer;
private bool locked = false;
public Buffer()
{
mutex = new Mutex(false);
buffer = new List<string>();
}
public bool isLocked()
{
return locked;
}
public void lockBuffer()
{
if (!locked)
{
//...
locked = true;
}
}
public void unlockBuffer()
{
if(locked)
{
mutex.ReleaseMutex();
locked = false;
}
}
public void wait()
{
mutex.WaitOne();
}
public void notify()
{
//...
}
}
It would be better if you use System.Collections.Concurrent.BlockingCollection. It doesn't require an external sync.
For those who don't use 4.0
using System;
using System.Collections.Generic;
using System.Threading;
namespace MyCollections
{
public class BlockingQueue<T> : IDisposable
{
Queue<T> _Queue = new Queue<T>();
SemaphoreSlim _ItemsInQueue = null;
SemaphoreSlim _FreeSlots = null;
int _MaxItems = -1;
public BlockingQueue(int maxItems=Int32.MaxValue)
{
_MaxItems = maxItems;
_ItemsInQueue = new SemaphoreSlim(0, maxItems);
_FreeSlots = new SemaphoreSlim(maxItems, maxItems);
}
public void Dispose()
{
if (_ItemsInQueue != null) _ItemsInQueue.Dispose();
if (_FreeSlots != null) _FreeSlots.Dispose();
}
public int Count
{
get { return _ItemsInQueue.CurrentCount; }
}
public void Add(T item)
{
if(_MaxItems != Int32.MaxValue) _FreeSlots.Wait();
lock (this)
{
_Queue.Enqueue(item);
_ItemsInQueue.Release();
}
}
public T Take()
{
T item = default(T);
_ItemsInQueue.Wait();
lock (this)
{
item = _Queue.Dequeue();
if (_MaxItems != Int32.MaxValue) _FreeSlots.Release();
}
return item;
}
}
}
The following code is not thread-safe. If two threads are entering this method at the same time, both might pass the if condition successfully.
public void lockBuffer()
{
if (!locked)
{
//...
locked = true;
}
}
You simply might want to do something like this:
lock (_sycnObject)
{
buffer.lockBuffer();
buffer.add(tcpPacket);
buffer.unlockBuffer();
buffer.notify();
}
I don't think you're doing something sophisticated that requires more than the simple to use lock-statement.
I wouldn't use Mutexes since I suppose you aren't dealing with multiple processes synchronization. Locks are pretty fine and simpler to implement:
class Buffer
{
private readonly object syncObject = new object();
private readonly List<string> buffer = new List<string>();
public void AddPacket(string packet)
{
lock (syncObject)
{
buffer.Add(packet);
}
}
public void Notify()
{
// Do something, if needed lock again here
// lock (syncObject)
// {
// Notify Implementation
// }
}
}
The usage is obviously (as you requested):
var myBuffer = new Buffer();
myBuffer.Add("Hello, World!");
myBuffer.Notify();
I'm attempting to make my simple C# graphics library multi-threaded. However, after the introduction of this code:
/* foreach (IAffector affector in affectorLookup.Values)
affector.Update(timestep); */
taskManager.Value = timestep; taskManager.Start();
foreach (IAffector affector in affectorLookup.Values)
taskManager.AddToQueue(affector.Update);
taskManager.StopWhenDone();
taskManager.Wait();
the simulation starts experiencing sharp lag-spikes, which seem to originate in TaskHandler.Run (I can't tell for sure, because adding the previous code makes my code profiler ignore anything outside TaskHandler.Run).
The task manager:
public class TaskManager
{
public delegate void MethodDel(float timestep);
private Queue<MethodDel> queue;
private List<TaskHandler> handlers;
private float value;
public float Value
{
get
{
return value;
}
set
{
this.value = value;
}
}
public TaskManager()
{
this.queue = new Queue<MethodDel>();
this.handlers = new List<TaskHandler>(System.Environment.ProcessorCount);
for (int t = 0; t < this.handlers.Capacity; ++t)
this.handlers.Add(new TaskHandler(this));
this.value = 0;
}
public void Start()
{
foreach (var handler in handlers)
handler.Start();
}
public void Stop()
{
lock (queue)
queue.Clear();
foreach (var handler in handlers)
handler.StopWhenDone();
}
public void StopWhenDone()
{
foreach (var handler in handlers)
handler.StopWhenDone();
}
public void AddToQueue(MethodDel method)
{
lock (queue)
queue.Enqueue(method);
}
public bool GetFromQueue(out MethodDel method)
{
lock (queue)
{
if (queue.Count == 0) { method = null; return false; }
method = queue.Dequeue();
return true;
}
}
public int GetQueueCount()
{
return queue.Count;
}
internal void Wait()
{
// Have to wait for them one at a time because the main thread is STA.
WaitHandle[] waitHandles = new WaitHandle[1];
// for (int t = 0; t < handlers.Count; ++t)
// waitHandles[t] = handlers[t].WaitHandle;
// WaitHandle.WaitAll(waitHandles);
for (int t = 0; t < handlers.Count; ++t)
{ waitHandles[0] = handlers[t].WaitHandle; WaitHandle.WaitAll(waitHandles); }
}
}
And the task handler:
public class TaskHandler
{
private TaskManager manager;
private Thread thread;
private bool stopWhenDone;
private ManualResetEvent waitHandle;
public ManualResetEvent WaitHandle
{
get
{
return waitHandle;
}
}
public TaskHandler(TaskManager manager)
{
this.manager = manager;
}
public void Start()
{
waitHandle = new ManualResetEvent(false);
stopWhenDone = false;
thread = new Thread(Run);
thread.IsBackground = true;
thread.SetApartmentState(ApartmentState.MTA);
thread.Start();
}
public void StopWhenDone()
{
this.stopWhenDone = true;
}
// Possible source of slowdown
private void Run()
{
TaskManager.MethodDel curMethod;
while (!stopWhenDone || manager.GetQueueCount() > 0)
{
if (manager.GetFromQueue(out curMethod))
{
curMethod(manager.Value);
}
}
waitHandle.Set();
}
}
Starting a thread is a heavy operation. Not sure if it's as heavy as you are experiencing, but that could be it. Also, having all your processing run parallel can be putting a big strain on your system with possibly little benefit...
I'm going to venture that the spikes have something to do with waitHandle.Set();
I like the overall design, but I have not used WaitHandle before, so I am unsure how this interacts with your design.