Introduction
Usually, it is very cumbersome to detect and locate memory leaks. This article will provide a way to locate memory leaks in .NET applications. First, I will talk about Resource Allocation and Garbage Collection algorithms, and then will talk about detecting leaks in .NET apps. Pay attention to the bold texts in the code section.
Background
Resource Allocation
The CLR (Common Language Runtime) allocates all the resources on the managed heap and releases them when they are no longer required by the application. C/C++ applications were prone to memory leaks because programmers had to manually allocate and free memory.
The Runtime maintains a NextObjPtr for the next free space on the heap. When a new process is initialized, the CLR allocates a contiguous space on the heap for this process represented by NextObjPtr, and increments the NextObjPtr pointer to the next free space. The space is contiguous, and is different from the C++ heap where it is maintained as a linked list. GC heap is efficient when compared to the C++ heap because to allocate new memory, the GC doesn't have to search from a list of free memory or through a linked list. As times pass by, gaps start to appear in the heap as objects get deleted, and so GC has to compact the heap, which is costly. GC in .NET uses the Win32 API VirtualAlloc or VirtualAllocEX to reserve memory.
.NET uses several types of memory like stack, unmanaged heap, and managed heap.
- Stack: It is managed on a per thread basis, and is used to store local variables, method parameters, and temporary values. GC doesn't clean the stack as its get automatically cleaned when the method returns. The references to objects are stored on the stack, but the actual object gets allocated on the heap and the GC is aware of that. When the GC cannot find a reference for an object, it removes it from the heap.
- Unmanaged Heap: Unmanaged code will allocate objects on the unmanaged heap or stack. Managed code can also allocate object on the unmanaged heap by calling Win32 APIs.
- Managed Heap: Managed code allocates objects on the managed heap, and the GC takes care of the management of the managed heap. The GC also maintains a Large Object Heap to compensate the cost of moving large objects in memory.
The garbage collector checks the heap for objects which are no longer used by the application. If such objects exist, then the GC removes those objects from the heap. Now, the question is how GC finds out about these objects which are not used by the application. Every application maintains a set of roots. Roots are like pointers to the objects on the heap. All global and static object pointers are considered as application roots. Any local variable on the thread stack is considered as application root. This list of roots is maintained by the JIT compiler and the CLR, and is made available to the GC.
When the GC starts running, it treats all objects as garbage, and makes an assumption that none of the objects on the heap are accessible. It then starts walking with the list of application roots, and starts building a graph of accessible objects. It marks all the objects on the heap as accessible if the objects are directly accessible lie an application root or indirectly accessible via any other object. For each application, the GC maintains a tree of references that tracks the objects referenced by the application. Using this approach, GC builds a list of live objects, and then walks through the heap in search of objects which are not present in this list of live objects. After finding out the objects which are not present in this list of live objects, it marks them all as garbage, and starts to compact the memory to remove holes which were created by unreferenced (dead) objects. It uses the memcpy function to move the objects from one memory location to another, and modifies the application roots to point to new locations.
If there is a live reference to the object, then it is said to be strongly rooted. .NET also has the concept of Weak Reference. Any object can be created as a weak reference which tells the GC that we want to access this object, but if the GC is going through garbage collection, then it can collect it. Weak reference is generally used for very large objects which are easy to create but are costly to maintain in memory.
Moving objects in memory gives a significant performance hit. To improve performance, the GC does several optimizations like large object heap and generations. Objects which are in size greater than 85,000 bytes are allocated on the large object heap. Moving large objects in memory is costly, so the GC maintains a separate heap for large objects, which it never compacts. The GC also maintains generations of objects. Whenever a new object is to be allocated and the managed heap doesn't have enough memory for the new object, a GC collection is performed. For the first time, every object in the heap is considered in Gen 0. After that, the GC performs a collection. The objects which survive are moved to Gen 1, and similarly, those which survive the Gen 1 collection move to Gen 2. The GC makes the assumption that a new object will have a short lifetime and an old object will have a longer lifetime. Whenever new memory is required, the GC tries to collect memory from Gen 0, and if enough memory can't be obtained from Gen 0 collections, then a Gen 1 or even Gen 2 collection is performed.
The following steps occur during each GC collection:
- Execution Engine Suspension – The EE is suspended until all managed threads have reached a point in their code execution deemed "safe".
- Mark – Objects that don't have roots are marked garbage.
- Plan – The GC creates a budget for each generation being collected, and then determines the amount of fragmentation that will exist in the managed heap as a result of a GC collection.
- Sweep – Deletes all objects marked for deletion.
- Compact – Moves all non-pinned objects that survived the GC to the lower end of the heap.
- Execution Engine Restart – Restart the execution of managed threads.
There are different kinds of roots in the GC:
- Strong Reference – If a strong reference exists for an object, then it is considered in use and is not collected during the next GC collection.
- Weak Reference – This is also a reference, but the object can live till the next GC collection. Weak Reference works like a cache for an object.
The GC can track an unmanaged resource's lifetime, but it can't reclaim the memory used by the resource unless destructors are used or code is written to override the Finalize in the base class.
A finalizer exists to allow the programmer to clean up the native resources used by the object before the object is garbage collected. But, using a finalizer promotes the object collection to the next generation. Whenever a new object with a Finalize method is allocated on the heap, a pointer to that object is placed on the Finalization queue. During garbage collection, if the GC finds that an object is not reachable, it then searches the Finalization queue for any reference to the object. If it finds a reference, then it removes the object from the Finalization queue and appends it to other data structure called a Freachable queue. At this point, the Garbage Collector has finished identifying garbage and compacts the memory. After that, the finalization thread empties the Freachable queue by executing each object's Finalize method. The next time a GC collection is performed, the GC sees this object as garbage and reclaims the memory assigned to this object.
It takes more time to reclaim the memory from objects having Finalize methods and affects performance, so a Finalize method should only be used when required.
Memory leaks can occur either in the stack, unmanaged heap, or the managed heap. There are many ways to find out that memory is leaking, like memory increasing in the Task Manager. Before starting to correct the memory problem, you need to determine the kind of memory which is leaking. Perfmon can be used to examine counters such as Process/Private bytes, .NET CLR Memory/# bytes in all heaps, and the .NET CLR LocksAndThreads/# of the current logical thread. If the .NET CLR LocksAndThreads/# is increasing unexpectedly, then the thread stack is leaking. If only Process/Private bytes are increasing but the .NET CLR Memory is not increasing, then unmanaged memory is leaking, else if both are increasing, then managed memory is leaking.
Figure 1 : Perfmon output for Privates Bytes and #Bytes in all heaps
Stack Memory
Stack memory gets reclaimed after the method returns. Stack memory can get leaked in two ways. First, a method call consumes a significant amount of stack resources that never returns, thereby never releasing the associated stack frame. The other is by creating background threads and never terminating them, thus leaking the thread stack.
Unmanaged Heap Memory
If the total memory usage is increasing but the .NET CLR memory is not increasing, then unmanaged memory is leaking. Unmanaged memory can leak in several ways - if the managed code is interoperating with unmanaged code and a leak exists in the unmanaged code. .NET doesn't make any guarantee that the finalizer for each object will get called. In the current implementation, .NET has one finalizer thread. If there exists a finalizer which blocks this thread, then the other finalizer will never get called and the unmanaged memory will leak which was supposed to be released. When an AppDomain is torn down, the CLR tries to run all the finalizers, but if a blocking finalizer exists, then it can prevent the CLR from completing the AppDomain tear down. To prevent this, the CLR implements a timeout on the process, after which it stops the finalization process, and the unmanaged memory which was supposed to be removed is left leaked.
Managed Heap Memory
Managed memory can also get leaked by several ways like fragmentation of the Large Object Heap. The memory in the Large Object Heap never gets compacted, so there is a loss in memory over there. Also, if there exist some objects which are not needed, but there exists a reference to the objects, then GC never claims the memory assigned to these objects.
This kind of leak is common, and can be resolved using SOS.dll. There are two ways to use SOS.dll:
- Open the application in Visual Studio 2005. Go to the properties of the project. In the Debug tab, make sure you have "Enable unmanaged code debugging" checked, or if you are attaching to a process, then in the "Attach to Process" window, click Select, and then select "Managed code" and "Native code". Run the application, and set a breakpoint in the code somewhere you want your application to break. Hit the breakpoint. Go to Debug -> Windows -> Immediate.
- Run .load SOS.dll
The SOS.dll is the most popular debugging extension used to debug managed code. It has many powerful commands that can obtain information such as managed call stack, details about managed heap, objects in heap, and much more.
- Run !dumpheap –stat or !dumpheap –type PolicyEditor
This command scans the GC heaps and lists the objects that lie therein. The –Stat argument is used to display the output to a statistical summary. More information about this (any) command can be found by using !help dumpheap.
The output will be like:
03f354ec 20 1120 View.Console.Configuration.Workspace
03713e44 35 1120 System.Windows.Forms.LayoutEventHandler
03710bec 56 1120 System.ComponentModel.Container+Site
024dc014 70 1120 System.Configuration.PropertySourceInfo
0492ed48 94 1128 Infragistics.Shared.SubObjectBase+NotifyId
081d9ac4 2 1136 View.Manger.UI.PolicyEditor
03d79898 1 1140 System.Text.RegularExpressions.RegexCharClass+
LowerCaseMapping[]
08857bdc 22 1144 Infragistics.Win.UltraWinToolbars.ComboBoxTool+
ComboBoxToolExtraSharedProps
0869c604 8 1152 Infragistics.Win.UltraWinEditors.UltraTextEditorUIElement
- The above command will list all the objects present in the memory. If you think that an object should not be present in memory and should have been garbage collected, then open a Find (Ctrl +F) window and type the name of the object and search for it. If you cannot find the object, then either it has been garbage collected, or it was not instantiated. If you find the object, for example, the
PolicyEditor object in the above list, then copy the MT (Method Table) address, which is 081d9ac4 in the above case. - Run !dumpheap -mt 081d9ac4
This will list all the objects with this MethodTable address. Those objects are the instances of PolicyEditor.
The output will be something like:
------------------------------
Heap 0
Address MT Size
total 0 objects
------------------------------
Heap 1
Address MT Size
total 0 objects
------------------------------
Heap 2
Address MT Size
28404fd8 081d9ac4 568
total 1 objects
------------------------------
Heap 3
Address MT Size
2c49f098 081d9ac4 568
total 1 objects
------------------------------
total 2 objects
Statistics:
MT Count TotalSize Class Name
081d9ac4 2 1136 View.Manger.UI.PolicyEditor
Total 2 objects
Copy the address of any instance, for example, 28404fd8.
- Run !gcroot 28404fd8
This will list the path to this object from the root of the GC tree. If some path exists for this object, then this is considered non garbage and is not collected by the GC, which could be a reason for memory leakage.
The output will be like:
Error during command: warning!
Extension is using a feature which Visual does not implement.
Scan Thread 6460 OSTHread 193c
Scan Thread 1884 OSTHread 75c
Scan Thread 7520 OSTHread 1d60
Scan Thread 7716 OSTHread 1e24
Scan Thread 0 OSTHread 0
Scan Thread 7428 OSTHread 1d04
Scan Thread 0 OSTHread 0
Scan Thread 4728 OSTHread 1278
DOMAIN(01D637C8):HANDLE(Pinned):22b148c:Root:302d9250(System.Object[])->
2908633c(System.EventHandler)->
29085f2c(System.Object[])->
2c53a4cc(System.EventHandler)->
2c53a340(Infragistics.Win.UltraWinStatusBar.UltraStatusBar)->
2c5513b0(System.Collections.Hashtable)->
2c5513e8(System.Collections.Hashtable+bucket[])->
2c5390ac(Infragistics.Win.Printing.UltraPrintPreviewControl)->
2c530668(Infragistics.Win.UltraWinGrid.UltraGridPrintDocument)->
2c530858(System.Drawing.Printing.PrintEventHandler)->
2c52eff0(View.Windows.InfragisticsControls.CommonUltraGrid)->
24487ed0(System.EventHandler)->
2c52df64(View.Manger.UI.PolicyClarificationBrowser)->
2448f478(System.EventHandler)->
2c4e3714(View.Manger.UI.PolicyEditor)
DOMAIN(01D637C8):HANDLE(Pinned):22c12f8:Root:302d52b8(System.Object[])->
282e427c(System.Collections.Generic.Dictionary`2[[System.Object, mscorlib],
[System.Collections.Generic.List`1
[[Microsoft.Win32.SystemEvents+SystemEventInvokeInfo, System]],
mscorlib]])->
282e4444(System.Collections.Generic.Dictionary`2+Entry[[System.Object, mscorlib],
[System.Collections.Generic.List`1
[[Microsoft.Win32.SystemEvents+SystemEventInvokeInfo, System]],
mscorlib]][])->
282ee76c(System.Collections.Generic.List`1
[[Microsoft.Win32.SystemEvents+SystemEventInvokeInfo, System]])->
2048e9c4(System.Object[])->
2c500534(Microsoft.Win32.SystemEvents+SystemEventInvokeInfo)->
2c500514(Microsoft.Win32.UserPreferenceChangedEventHandler)->
2c4fff54(Infragistics.Win.UltraWinStatusBar.UltraStatusBar)->
2446c824(System.Collections.Hashtable)->
2446c85c(System.Collections.Hashtable+bucket[])->
2c4fecc0(Infragistics.Win.Printing.UltraPrintPreviewControl)->
2c4f6a24(Infragistics.Win.UltraWinGrid.UltraGridPrintDocument)->
2c4f6c14(System.Drawing.Printing.PrintEventHandler)->
2c4f53ac(View.Windows.InfragisticsControls.CommonUltraGrid)->
2446e90c(System.EventHandler)->
24450e80(View.Manger.UI.EvidenceDefinitionMiniBrowser)->
2c4e34bc(View.Manger.Views.PolicyEditController)->
2c4e3524(View.Manger.Views.PolicyEdit)->
2449cfe0(System.Windows.Forms.LayoutEventArgs)->
2c4e3714(View.Manger.UI.PolicyEditor)
DOMAIN(01D637C8):HANDLE(WeakLn):22c1d88:Root:24496d2c(
System.Windows.Forms.NativeMethods+WndProc)->
2c4e39fc(System.Windows.Forms.Control+ControlNativeWindow)->
2c4e3714(View.Manger.UI.PolicyEditor)
DOMAIN(01D637C8):HANDLE(WeakSh):22e313c:Root:2c4e39fc(
System.Windows.Forms.Control+ControlNativeWindow)
- In the above output, we can see that the
PolicyClarificationBrowser object has an event handler which holds a reference to PolicyEditor. To find out who is hooking up this event, take the address of the event handler, which is 2448f478, and dump this object. - !dumpobj 2448f478
Name: System.EventHandler
MethodTable: 7910d61c
EEClass: 790c3a7c
Size: 32(0x20) bytes
(C:\WINDOWS\assembly\GAC_32\mscorlib\2.0.0.0__b77a5c561934e089\mscorlib.dll)
Fields:
MT Field Offset Type VT Attr Value Name
790f9c18 40000f9 4 System.Object 0 instance 2c4e3714 _target
79109208 40000fa 8 ...ection.MethodBase 0 instance 00000000 _methodBase
790fe160 40000fb c System.IntPtr 0 instance 140144060 _methodPtr
790fe160 40000fc 10 System.IntPtr 0 instance 0 _methodPtrAux
790f9c18 4000106 14 System.Object 0 instance 00000000 _invocationList
790fe160 4000107 18 System.IntPtr 0 instance 0 _invocationCount
- If you take the address of the target in the above output, which is 2c4e3714, and run a !dumpobj with this address, then you can see in the output that it is of type
PolicyEditor. To get the method that is hooked up as a handler, convert the int value in _methodPtr to hex.
?0n140144060
(This command will not work in the Immediate Window. To run this command, you need to attach WinDbg to this or any other managed process, or use Google to do that.)
The output will be:
Evaluate expression: 140144060 = 085a6dbc
And then run !ip2md 085a6dbc
Failed to request MethodData, not in JIT code range
Sometimes this works, and will give you the name of the method, but if doesn't work, then we can dump the object and then find out the method.
- !dumpobj 2c4e3714
The output will be:
Name: View.Manger.UI.PolicyEditor
MethodTable: 0820865c
EEClass: 08219fd8
Size: 568(0x238) bytes
(C:\Documents and Settings\testuser\Application Data\View\SMC-D-44725-B\View.Manger.UI.dll)
Fields:
MT Field Offset Type VT Attr Value Name
790f9c18 4000184 4 System.Object 0 instance 00000000 __identity
024c1798 40008bc 8 ...ponentModel.ISite 0 instance 00000000 site
036f9c3c 40008bd c ....EventHandlerList 0 instance 2448f8ac events
790f9c18 40008bb 108 System.Object 0 static 242e1afc EventDisposed
- !dumpmt -md 0820865c
The output will be:
EEClass: 08219fd8
Module: 036e48d8
Name: View.Manger.UI.PolicyEditor
mdToken: 02000069 (C:\Documents and Settings\testuser\Application Data\
View\SMC-D-44725-B\View.Manger.UI.dll)
BaseSize: 0x238
ComponentSize: 0x0
Number of IFaces in IFaceMap: 15
Slots in VTable: 430
--------------------------------------
MethodDesc Table
Entry MethodDesc JIT Name
022a9fc5 022a9ec8 NONE System.ComponentModel.Component.ToString()
793539c0 7913bd50 PreJIT System.Object.Equals(System.Object)
…
…
085a6d44 08208578 NONE View.Manger.UI.PolicyEditor.btnStatementDelete_Click(
System.Object, System.EventArgs)
085a6de4 08208580 NONE View.Manger.UI.PolicyEditor.btnProperties_Click(
System.Object, System.EventArgs)
085a6dbc 08208588 NONE View.Manger.UI.PolicyEditor.ClarificationBrowser_OpenClick(
System.Object, System.EventArgs)
08208f0d 08208590 NONE View.Manger.UI.PolicyEditor.OpenClarification()
085a6dd0 08208598 NONE View.Manger.UI.PolicyEditor.Clarification_SelectionChanged(
System.Object, View.Manger.UI.SelectEventArgs)
085a6d94 082085a0 NONE View.Manger.UI.PolicyEditor.btnExceptionNew_Click(
System.Object, System.EventArgs)
…
…
In the above output, search for the hex address we calculated above from _methodptr, and you will get the name of the method which is still holding a reference to the PolicyEditor object and is the reason for the memory leak.
History
- Initially posted on Jul 3, 2007.
