Jon Cole's WebLog : Debugging

Creating a Process Memory Dump

joncole — Fri, 30 Mar 2007 07:29:00 GMT

When trying to help customers debug some issues it is periodically necessary to get a memory dump (snapshot) of a process for analysis at another time or place. This is frequently the case when an error is intermittent or a live debug session is not feasible (like in a production environment). I will discuss the steps necessary to install the right debuggers, attach to the process in question, looking for the right event to occur (knowing when to save the dump) and saving a dump to disk. In this case I will be looking at BtsNtSvc.exe, which is the BizTalk NT Service process.

Installing the Debuggers

Many debuggers have support for creating (or opening) memory dumps, including Visual Studio 2005. However, because of the size of Visual Studio and license requirements, I will discuss the use of the freely downloadable debugging tools that Microsoft provides - Debugging Tools for Windows. Make sure you install the correct version for you processor architecture (x86 vs x64).

You may want to read the Getting Started page for supplemental information. These tools come with several debuggers, most of which support the same types of operations. Windbg.exe is the GUI version, but you can also use cdb.exe (a command line based debugger) and accomplish pretty much the same things because they are built on top of the same underlying debugging libraries. In this article I will focus on WinDbg.

Setting your Symbol Path

The symbol path (path to pdb files) can be set in several ways. You can set it before you start the debugger by using the _NT_SYMBOL_PATH environment variable or at runtime. To set the symbol path at runtime, select File->Symbol File Path. This symbol path is basically just a semicolon (;) delimited list of search paths that the debugger should use when trying to locate debug symbols for the exe/dlls in your process. I will try to write more about setting up your symbol path at a later date.

You can also set the symbol path by using the .sympath command (more information below on how to run commands). See the debugger help files for more information.

Check out KB article 311503 for more information on setting symbols, including how to get publicly available symbols for microsoft products. It aslo includes a video tuturial on how to set the symbol path in the Visual Studio.Net debugger.

Attaching to the process

You can attach to the process in a couple of ways. Choose one of the following:

From the command line you can type any of the following:
1. "Windbg.exe -p <processID>" (example: "windbg.exe -p 1234")
2. "windbg.exe -pn <processName>" (example: "windbg.exe -pn btsntsvc.exe")
3. Note that you can add "-g" (lower case g) to any of the above to avoid breaking into the debugger immediately after attaching.
After starting WinDbg:

Select File->Attach to a Process from the menu
Select the Process from the list in the new windows that comes up

Once you have the debugger open, you can either type ".hh" in the command box or by selecting Help-> Contents from the menu. Note that you will have to break into the debugger before you can run any commands in the command box. Hit "Ctrl-Break" on your keyboard or click the pause button on the toolbar to break into the debugger.

FYI, here is a snapshot of the command box (found in the lower left hand corner of the window). I have circled it in red. Note that the "0 : 003" indicates that I am currently examining thread number 3.

Looking for the right event

There may be several types of events you are interested in and several ways to locate the event you are interested in. I will only speak of three types of events in this article. Some will result in the debugger breaking for you automatically, others you will have to tell the debugger you are interested in stopping if it occurs. I will primarily focus on .NET exceptions in this article.

Access Violation (a.k.a. AV). This type of error will result in the process breaking into the debugger without any special setup on your part. AVs occur when your program tries to access memory that it doesn't have permission to access. Examples would be reading/writing protected memory or trying to use an object that is null. In .NET, these will result in either an AccessViolationException (read/write protected memory) or NullReferenceException (null object referenced in code).
.NET exception at the time it is thrown (a.k.a first chance exception). Windbg will not break into the debugger for this one automatically. You have to tell it that you want it to stop. You can do this by running the command sxe clr. If you need to stop breaking into the debugger every time a .NET exception is thrown, you can use the sxi clr command. Keep in mind that just because an exception is thrown does not mean that it is unexpected and unhandled. The process may expect to get exceptions in certain recoverable situations and the user may be completely unaware that anything has happened.
.NET exception is thrown and is not handled. This will result in your process crashing. This is especially true in .NET 2.0 and later. In earlier versions of the framework, an exception being thrown and unhandled on a thread pool thread did not necessarily cause the application to shut down - it was swallowed. This made debugging applications more difficult, so in V2.0 of the framework an unhandled exception causes the application to shutdown immediately to make the system easier to debug.

Now that we know how to get the system to break into the debugger, we need to be able to look at the exception that occurred to determine if it is the one we care about. I will discuss how to do this in V2.0 of the framework. It can be done in earlier versions of the framework also, but it isn't quite as easy.

Because Windbg is primarily a native application debugging tool, you have to load what is called a debugger extension to enable you to debug managed code. This extension is called sos.dll and it sits in the .net v2.0 installation dir (same dir as mscorwks.dll). You can use the following command to easily load it

.loadby sos mscorwks

This tells windbg to look in the same directory as mscorwks.dll (always loaded into .NET applications) to find sos.dll. Command that are loaded from a debugger extension dll are always prefixed with an exclamation mark (!).

When you have sos.dll loaded you can run !help. Note that if you have multiple extensions loaded, you may have to prefix your commands with the specific extension that has the command you want. For example: !sos.help.

Now, when the debugger breaks, we can do run two commands to help us figure out where we are and what type of exception occured. !sos.printexception (or !pe is the shorthand version) and !clrstack. If the exception type, message and stack trace match the one you see in logs, event viewer, etc., then you have (hopefully) found the right exception.

Here is some sample output when a first chance .NET exception occurs:

(650.1c9c): CLR exception - code e0434f4d (first chance)
First chance exceptions are reported before any exception handling.
This exception may be expected and handled.
eax=0012f298 ebx=00181a60 ecx=00000000 edx=00000025 esi=0012f324 edi=e0434f4d
eip=77e55e02 esp=0012f294 ebp=0012f2e8 iopl=0 nv up ei pl nz na po nc
cs=001b ss=0023 ds=0023 es=0023 fs=003b gs=0000 efl=00000202
KERNEL32!RaiseException+0x53:
77e55e02 5e pop esi

0:000> !pe
Exception object: 01355858
Exception type: System.InvalidOperationException
Message: This is a sample exception message for an InvalidOperationException
InnerException: <none>
StackTrace (generated):
<none>
StackTraceString: <none>
HResult: 80131509

0:000> !clrstack
OS Thread Id: 0x1c9c (0)
ESP EIP
0012f370 77e55e02 [HelperMethodFrame: 0012f370]
0012f414 00d8021e Test.AnotherFunctionThatThrowsAnException(System.String)
0012f420 00d801c2 Test.SomeFunction()
0012f428 00d800b2 Test.Main()
0012f69c 79e88f63 [GCFrame: 0012f69c]

If the exception you are currently looking at is not the one you were looking for, then you can tell the debugger to let the process continue until the next event by using the g command (go). Of course, these debuggers support break points also, but that is a discussion for another day.

Saving the Dump File

Now that you have the process broken into the debugger at the right place, you can use .dump command to write the memory dump out to a file.

.dump <options> <path to file>
example: .dump /mfh c:\temp\myDump.dmp

I would recommend the /mfh options for this command as I show above. You can read the help files for the .dump command to get more information on the options you can use and what they imply.

Just a reminder...

One last warning about memory dumps before we are done. Memory dumps contain the memory snapshot of the process. This memory can contain private information that you and/or your customers wouldn't want to give out to just anyone. Some examples: usernames and passwords, credit card numbers, etc - anything that is in memory at the time the dump is taken. Consider yourself warned - don't give a copy of the memory dump to just anyone.

Debugging a memory leak in managed code: Ping - SendAsync

joncole — Fri, 16 Dec 2005 02:13:00 GMT

Included in version 2.0 of the .NET framework is a Ping class (System.Net.NetworkInformation namespace) that can be used to monitor the up/down status of a network connection to a server. We recently realized that we had a minor bug in the class that can present itself as a major problem - a memory leak. Fortunately there is an easy workaround to this problem. I thought I would take this as an oportunity to demonstrate how to use the windbg debugger as I explain how we figured out what was going wrong.

Lets consider this simple console ping application that sends a series of ping requests to a server using the SendAsync method.

using System;
using System.Threading;
using System.Net.NetworkInformation;

class program {
static int PingsOutstanding;

static void Main(string[] args)
{
  string targetServer = "www.contoso.com";
  for (int i = 0; i < 1000; i++) {
   PingServer(targetServer);
  }

  int currentcount = PingsOutstanding;
  while (currentcount > 0) {
   Console.WriteLine("waiting for {0} ping requests to complete", currentcount);
   Thread.Sleep(1000);
   currentcount = PingsOutstanding;
  }

  //the calls to the GC are just for debugging purposes
  Console.WriteLine("Calling GC.Collect()");
  GC.Collect();
  GC.WaitForPendingFinalizers();

Console.WriteLine("Press <enter> to exit application");
Console.ReadLine();
}

static void PingServer(string hostnameoraddress)
{
  try {
   Ping pingsender = new Ping();
   pingsender.PingCompleted += new PingCompletedEventHandler(OnPingCompleted);
   pingsender.SendAsync(hostnameoraddress, null);
   Interlocked.Increment(ref PingsOutstanding);
  } catch (Exception ex) {
   Console.WriteLine(ex);
  }
}

static void OnPingCompleted(object sender, PingCompletedEventArgs e)
{
Interlocked.Decrement(ref PingsOutstanding);

try {
Ping pingsender = (Ping)sender;

   pingsender.Dispose();
  } catch (Exception ex) {
   Console.WriteLine(ex);
  }
}
}

This application will not consume a ton of memory under normal operation but if you increase the number of ping requests that are sent, you will see that memory consumption also climbs. Given that the .Net framework uses a garbage collection approach to memory allocation and cleanup, you would expect that the memory would be cleaned up properly within a reasonable amount of time. You should also notice that I am calling pingSender.Dispose() which we would hope would release the resources being held by this object.

For demonstration purposes I have added a Console.ReadLine() call at the end of the application so that it won't exit. I have also added a call to GC.Collect() and GC.WaitForPendingFinalizers() to make sure that when I go to debug this I am certain that the GC has recently run and cleaned up any objects that are available for cleanup. It is not usually advisable to force GC collections except for debugging purposes. See Maoni's blog for great information on how the GC works and how to use it effectively.

Using Windbg.exe

I chose to use Windbg because I have found it to be the best debugger for memory leaks. It is a very powerful GUI debugger that is used very frequently inside of Microsoft for debugging everything from application bugs to OS bugs. You can download it from here: http://www.microsoft.com/whdc/devtools/debugging/default.mspx.

A couple of quick notes before I begin:

updates to these debugging tools are available regularly, so make sure to update periodically.
I will color commands that are to be run inside the debugger in green.
I have access to all of the debugger symbols for the .Net framework, but you don't. However, most of the information you need should be accessible by using ".symfix" as noted below. The principles discussed here apply to debugging a memory leak in any managed application.
- public microsoft symbols can loaded into windbg using the .symfix followed by a .reload command (requires internet connection to a microsoft http server)
I may trim sample output (colored in purple) from the debugger in order to focus on the important information.

Now, Windbg is a native windows debugger and doesn't really know much about the .Net framework, so you have to load an extension to the debugger that understands managed applications. SOS.dll is the debugger extension we are looking for. In the command window this will load the sos extension:

.loadby sos mscorwks

This command tells the debugger to load the sos dll by looking in the same directory as the mscorwks dll, which is a core dll that is always loaded in a managed application. Note that Windbg breaks into the debugger immediately upon loading the application, so mscorwks will not be loaded at this point. You will have to run the application long enough for the .NET framework code to be loaded. You can actually tell the debugger to break when mscorwks is loaded by running this in the command window:

sxe ld:mscorwks

There are two ways that you can run commands from the sos this extension dll:

!<command> or !sos.<command>

Here are examples for the help function found in sos (it prints usage information):

!help
!sos.help

The second one is useful when you have multiple extensions that share common function names.

When debugging this leak, I actually let the app run until it blocked on the call to Console.ReadLine() so that GC should have cleaned up. The first thing I do when dealing with a managed memory leak is to look at all the objects that have been created by the runtime and see if I see any that jump out at me. Lets look at the heap and see what it contains.

   !dumpheap -stat
total 20600 objects
Statistics:
      MT    Count    TotalSize Class Name
7a76eb14        1           12 System.Net.DnsPermission
7a755834        1           12 System.Diagnostics.PerformanceCounterCategoryType
7a753394        1           12 System.Diagnostics.TraceOptions
7a71a710        1           12 System.Net.TimeoutValidator
7912b908        1           12 System.Collections.Generic.GenericEqualityComparer`1[[System.String, mscorlib]]
79112c90        1           12 System.Collections.Comparer
7910db30        1           12 System.Threading.SynchronizationContext
7910a718        1           12 System.DefaultBinder
79107f40        1           12 System.RuntimeTypeHandle
79107ac4        1           12 System.Reflection.__Filters
79102f48        1           12 System.__Filters
79102ef8        1           12 System.Reflection.Missing
79101ca8        1           12 System.RuntimeType+TypeCacheQueue
79100800        1           12 System.Text.DecoderExceptionFallback
...(output trimmed)
0105a514       77         4004 System.Configuration.FactoryRecord
791242ec       47         8136 System.Collections.Hashtable+bucket[]
79160a3c     1000        16000 System.Threading.RegisteredWaitHandle
79124228       58        19216 System.Object[]
7a779154     1000        20000 System.Net.SafeLocalFree
7a778ec0     1000        20000 System.Net.SafeCloseHandle
7a761bb0     1000        20000 System.ComponentModel.AsyncOperation
79160b84     1000        20000 System.Threading._ThreadPoolWaitOrTimerCallback
7910cf3c     1001        20020 Microsoft.Win32.SafeHandles.SafeWaitHandle
791609c8     1000        24000 System.Threading.RegisteredWaitHandleSafe
79115d0c     1000        24000 System.Threading.ManualResetEvent
7a7811f4     1000        32000 System.Net.NetworkInformation.PingCompletedEventHandler
79160aa8     1000        32000 System.Threading.WaitOrTimerCallback
7915ff38     1000        32000 System.Threading.SendOrPostCallback
7910d2f4     1001        36036 System.Threading.ExecutionContext
79124418     1004        44820 System.Byte[]
7a7812e0     1000        88000 System.Net.NetworkInformation.Ping
00150c90      421       108132      Free
790fa3e0     5262       304552 System.String
Total 20600 objects

Looking at the columns in the output, MT stands for Method Table and is basically a pointer to the table that describes that type of object. Count is the number of objects that exist in the heap of the given type. TotalSize is the amount of memory that is being consumed by any one type of object. The last column is obviously the fully typed name of the object.

The first thing that jumps out at me in this case is the fact that we still have a large number of ping objects in the heap even though we aren't holding onto any references in our code. We also called dispose on these objects, so this is a huge red light telling me that something went wrong. Also, the fact that we sent exacly 1000 ping requests and there are exactly 1000 ping objects still hanging around calls my attention. Lets take a look at these ping objects and see why they aren't getting cleaned up (Below I show two ways of doing this, you really only need to do one of them).

!dumpheap -type Ping

This does a substring match on any object with Ping in it. In this case it will list both System.Net.NetworkInformation.Ping objects and System.Net.NetworkInformation.PingCompletedEventHandler objects. I really wanted just the Ping objects, so we could have run this command (using the MT for the ping object - obtained above)

!dumpheap -mt 7a7812e0

Address       MT     Size
01271ff0 7a7812e0       88
012720f8 7a7812e0       88
012721d0 7a7812e0       88
012722a8 7a7812e0       88
01272380 7a7812e0       88
01272458 7a7812e0       88
01272530 7a7812e0       88
01272608 7a7812e0       88
012726e0 7a7812e0       88
012727b8 7a7812e0       88
01272890 7a7812e0       88
01272968 7a7812e0       88
...(output trimmed)
012fada0 7a7812e0       88
012faeec 7a7812e0       88
012fb038 7a7812e0       88
012fb184 7a7812e0       88
012fb2d0 7a7812e0       88
012fb41c 7a7812e0       88
012fb568 7a7812e0       88
012fb6b4 7a7812e0       88
012fb800 7a7812e0       88
012fb94c 7a7812e0       88
012fba98 7a7812e0       88
012fbbe4 7a7812e0       88
012fbd30 7a7812e0       88
012fbe7c 7a7812e0       88
total 1000 objects
Statistics:
      MT    Count    TotalSize Class Name
7a7812e0     1000        88000 System.Net.NetworkInformation.Ping
Total 1000 objects

Notice that the MT for all of the above objects matches that for the ping object. The left hand column is the memory address where the object resides. We can use this address to examine the actual ping object. To take a look at the first object in the above list we would use this command:

!dumpobject 01271ff0

I am not going to go into details at this point because this won't lead us where the problem lies. I thought I would list the dumpobject command just to point it out to you. The problem in this case is that objects that we think should be cleaned up are not cleaned up as expected. This means that there must be a reference to the object somewhere. The following command will help you to see what objects hold references to an object in question. We will use the same object address we used in the dumpobject example:

!gcroot 01271ff0

Note: Roots found on stacks may be false positives. Run "!help gcroot" for
more info.
Scan Thread 0 OSTHread 2f8
Scan Thread 2 OSTHread fcc
Scan Thread 3 OSTHread de8
Scan Thread 4 OSTHread eb8
Scan Thread 8 OSTHread c8c
Scan Thread 12 OSTHread 464
Scan Thread 13 OSTHread 1e0
DOMAIN(0014BF60):HANDLE(Strong):8f15b8:Root:0128d2c0(System.Threading._ThreadPoolWaitOrTimerCallback)->
01271ff0(System.Net.NetworkInformation.Ping)

This tool scans the threads to find any object that holds a reference to the given memory address. We can see that some threadpool object is holding onto us. If we go back to the objects on the heap we can see that there are these _ThreadPoolWaitOrTimerCallback objects on the heap also. Now, I don't know a ton about the thread pool, but I would not expect those callback objects to still be on the heap either because they are most likely specific to a single usage of a threadpool thread. We could dump all of the _ThreadPoolWaitOrTimerCallback objects and check their roots (the objects holding onto them), but why dump the more objects from the heap? The output from !gcroot gave us a pointer to an instance that we can quickly examine.

!gcroot 0128d2c0

We would have hoped that the Dispose() function would have cleaned these objects up, but it didn't. So, what about the GC? Why didn't GC take care of them when dispose didn't? Notice how this gave us the same exact output from the gcroot of the ping class. This must mean that we have a circular reference (ping has a reference to this thread pool object and visa versa) and that is why the GC didn't clean it up.

Workaround (and reason why calling Dispose didn't clean this up)

The simple workaround to this problem is to cast the object to the IDisposable interface and then call Dispose on that casted object.

((IDisposable)pingSender).Dispose();

Why does this solve the problem? Take a look at the declaration of the class:

public class Ping:Component,IDisposable

Notice that it inherits from Component and it also implementes the IDisposable interface. Now lets look at the function declaration for Dispose on the Ping class:

void IDisposable.Dispose ()

In the implementation, the Dispose function was defined as part of the IDisposable interface but the fact that the base class Component also implements a Dispose function was missed. The Ping class doesn't override the Component.Dispose method and so if you call Dispose directly on the Ping object without casting it to an IDisposable Interface, the runtime calls the base class implementation of Dispose instead of the one that was intended (the one for IDisposable).

Note: There is another great blog on debugging managed memory leaks here: http://blogs.msdn.com/mvstanton/archive/2005/10/11/479861.aspx