20 November 2012

Understand and Prevent Deadlocks


Can you explain a typical C# deadlock in a few words? Do you know the simple rules that help you to write deadlock free code? Yes? Then stop reading and do something more useful.

If several threads have read/write access to the same data it is often often necessary to limit access to only on thread. This can be done with C# lock statement. Only one thread can execute code that is protected by a lock statement and a lock object. It is important to understand that not the lock statement protects the code, but the object given as an argument to the lock statement. If you don't know how the lock statement works, please read the msdn documentation before continuing. Using a lock statement is better than directly using a Mutex or EventWaitHandle because it protects you from stale locks that can occur if you forget to release your lock when an exception happens.

A deadlock can occur only if you use more than one lock object and the locks are acquired by each thread in a different order. Look at the following sequence diagram:



There are two threads A and B and two resources X and Y. Each resource is protected by a lock object.
Thread A acquires a lock for Resource X and continues. Then Thread B acquires a lock for Y and continues. Now Thread A tries to acquire a lock for Y. But Y is already locked by Thread B. This means Thread A is blocked now and waits until Y is released. Meanwhile Thread B continues and now needs a lock for X. But X is already locked by Thread B. Now Thread A is waiting for Thread B and Thread B is waiting for Thread A both threads will wait forever. Deadlock!

The corresponding code could look like this.

public class Deadlock
{
    static readonly object X = new object();
    static readonly object Y = new object();
   
    public void ThreadA()
    {
        lock(X)
        {           
            lock(Y)
            {
                // do something
            }
        }
    }

    public void ThreadB()
    {
        lock(Y)
        {
            lock(X)
            {
                // do something
            }
        }
    }
}

Normally nobody will write code as above with obvious deadlocks. But look at the following code, which is deadlock free:


public class Deadlock
{
    static readonly object X = new object();
    static readonly object Y = new object();
    static object _resourceX;
    static object _resourceY;

    public object ResourceX
    {
        get { lock (X) return _resourceX; }
    }

    public object ResourceY
    {
        get
        {
            lock (Y)
            {
                return _resourceY ?? (_resourceY = "Y");
            }
        }
    }

    public void ThreadA()
    {
        Console.WriteLine(ResourceX);
    }

    public void ThreadB()
    {
        lock(Y)
        {
            _resourceY = "TEST";
            Console.WriteLine(ResourceX);
        }
    }
}


But after re-factoring the getter for ResourceX to this

get { lock (X) return _resourceX ?? ResourceY; }

you have the same deadlock as in the first code sample!

Deadlock prevention rules


  1. Don't use static fields. Without static fields there is no need for locks.
  2. Don't reinvent the wheel. Use thread safe data structures from System.Collections.Concurrent or System.Threading.Interlocked before pouring lock statements over your code.
  3. A lock statement must be short in code and time. The lock should last nanoseconds not milliseconds.
  4. Don't call external code inside a lock block. Try to move this code outside the lock block. Only the manipulation of known private data should be protected. You don't know if external code contains locks now or in future (think of refactoring).

If you are following these rules you have a good chance to never introduce a deadlock in your whole career.



14 November 2012

How to find deadlocks in an ASP.NET application almost automatically


This is a quick how-to for finding deadlocks in an IIS/ASP.NET application running on a production server with .NET4 or .NET 4.5.

A deadlock bug inside your ASP.NET application is very ugly. And if it manifests only on some random production server of your web farm, maybe you feel like doom is immediately ahead. But with some simple tools you can catch and analyze those bugs.

These are the tools you need:
  • ProcDump from SysInternals   
  • WinDbg from Microsoft (Available as part of the Windows SDK (here is even more info)) 
  • sos.dll (part of the .NET framework)
  • sosext from STEVE'S TECHSPOT (Copy it into your WinDbg binaries folder)
ProcDump will be needed on the server where the deadlock occurs. All the other tools are only needed on your developer machine. Because WinDbg doesn't need any installation you can also prepare an emergency USB stick (or file share) with all the necessary tools.

If you think a deadlock occurred do the following:
  1. Connect to the Server
  2. Open IIS Manager 
  3. Open Worker Processes 
  4. Select the application pool that is suspected to be deadlocked
  5. Verify that you indeed have a deadlock, see the screenshot below
  6. Notice the <Process-ID> (see screenshot)
  7. Create a dump with procdump <Process-ID> -ma
    There are other tools, like Task Manager or Process Explorer, that could dump but only ProcDump is smart enough to create 32bit dumps for 32bit processes on a 64bit OS.
  8. Copy the dump and any available .pdb (symbol) files to your developer machine. 
  9. Depending on the bitness of your dump start either WinDbg (X86) or WinDbg (X64)
  10. Init the symbol path (File->Symbol File Path ...)
    SRV*c:\temp\symbols*http://msdl.microsoft.com/download/symbols
  11. File->Open Crash Dump
  12. Enter the following commands in the WinDbg Command Prompt and wait
  13. .loadby sos clr
  14. !load sosex
  15. !dlk
You should now see something like this:

0:000> .loadby sos clr
0:000> !load sosex
This dump has no SOSEX heap index.
The heap index makes searching for references and roots much faster.
To create a heap index, run !bhi
0:000> !dlk
Examining SyncBlocks...
Scanning for ReaderWriterLock instances...
Scanning for holders of ReaderWriterLock locks...
Scanning for ReaderWriterLockSlim instances...
Scanning for holders of ReaderWriterLockSlim locks...
Examining CriticalSections...
Scanning for threads waiting on SyncBlocks...
*** WARNING: Unable to verify checksum for mscorlib.ni.dll
Scanning for threads waiting on ReaderWriterLock locks...
Scanning for threads waiting on ReaderWriterLocksSlim locks...
*** WARNING: Unable to verify checksum for System.Web.Mvc.ni.dll
*** ERROR: Module load completed but symbols could not be loaded for System.Web.Mvc.ni.dll
Scanning for threads waiting on CriticalSections...
*** WARNING: Unable to verify checksum for System.Web.ni.dll
*DEADLOCK DETECTED*
CLR thread 0x5 holds the lock on SyncBlock 0126fa70 OBJ:103a5878[System.Object]
...and is waiting for the lock on SyncBlock 0126fb0c OBJ:103a58d0[System.Object]
CLR thread 0xa holds the lock on SyncBlock 0126fb0c OBJ:103a58d0[System.Object]
...and is waiting for the lock on SyncBlock 0126fa70 OBJ:103a5878[System.Object]
CLR Thread 0x5 is waiting at System.Threading.Monitor.Enter(System.Object, Boolean ByRef)(+0x17 Native)
CLR Thread 0xa is waiting at System.Threading.Monitor.Enter(System.Object, Boolean ByRef)(+0x17 Native)


1 deadlock detected.


Now you know that the managed threads 0x5 and 0xa are waiting on each other. With the !threads command you get a list of all threads. The Id Column (in decimal) is the managed thread id. To the left the WinDbg number is written. With  ~[5]e!clrstack command you can see the stacktrace of CLR thread 0x5. Or just use ~e*!clrstack to see all stacktraces. With this information you should immediately see the reason for the deadlock and start fixing the problem..


Deadlocked Requests visible in IIS Worker Process

Automate the Deadlock Detection

If you are smart, create a little script that automates step 2 to 7. We use this powershell script for checking every minute for a deadlock situation:


param($elapsedTimeThreshold, $requestCountThreshold)
Import-Module WebAd*
$i = 1
$appPools = Get-Item IIS:\AppPools\*
while ($i -le 5) {
ForEach ($appPool in $appPools){
 $count = ($appPool | Get-WebRequest | ? { $_.timeElapsed -gt $elapsedTimeThreshold }).count
 if($count -gt $requestCountThreshold){
$id = dir IIS:\AppPools\$($appPool.Name)\WorkerProcesses\ | Select-Object -expand processId
$filename = "id_" +$id +".dmp"
$options = "-ma"
$allArgs = @($options,$id, $filename)
procdump.exe $allArgs
 }
}
Start-Sleep -s 60
}


07 November 2012

Cooperative Thread Abort in .NET


Did you know that .NET uses a cooperative thread abort mechanism?

As someone coming from a C++ background I always thought that killing a thread is bad behavior and should be prevented at all costs. When you terminate a Win32 thread it can interrupt the thread in any state at every machine instruction so it may leave corrupted data structures behind.
A .NET thread cannot be terminated. Instead you can abort it. This is not just a naming issue, it is indeed a different behavior. If you call Abort() on a thread it will throw a ThreadAbortException in the aborted thread. All active catch and finally blocks will be executed before the thread gets terminated eventually. In theory, this allows the thread to do a proper cleanup. In reality this works only if every line of code is programmed in a way that it can handle a ThreadAbortException
in a proper way. And the first time you call 3rd-party code not under your control you are doomed.

Too make the situation more complex there are situations where throwing the ThreadAbortException is delayed. In versions prior to .NET 4.5 this was poorly documented but in the brand new documentation of the Thread.Abort method is very explicit about this. A thread cannot be aborted inside a catch or finally block or a static constructor (and probably not the initialization of static fields also) or any other constrained execution region.


Why is this important to you?

Well, if you are working in an ASP.NET/IIS environment the framework itself will call Abort() on threads which are executing too long. In this way the IIS can heal itself if some requests hit bad blocking code like endless loops, deadlocks or waiting on external requests. But if you were unlucky enough to implement your blocking code inside static constructors, catch or finally blocks your requests will hang forever in your worker process. It will look like the httpRuntime executionTimeout is not working and only a iisreset will cure the situation.

Download the CooperativeThreadAbortDemo sample application.