Finding Memory Leaks in Windows 2000

Finding Memory Leaks in Windows 2000

How to Use Umdh.exe to Find Memory Leaks This information applies to:

Microsoft Windows 2000 Advanced Server Microsoft Windows 2000 Server Microsoft Windows 2000 Professional Microsoft Windows NT Server version 4.0

Microsoft Windows NT Workstation version 4.0

User-Mode Dump Heap Utility (UMDH)
The user-mode dump heap utility works with the operating system to analyze the Windows heap allocations for a specific process. While this utility, and the other tools associated with it, are primarily targeted for Windows 2000, a Umdh_nt4.exe application, and the Dbghelp.dll which is not available on Windows NT, have been supplied for use under Windows NT version 4.0 Service Pack 6 (SP6). The following discusses how you can use UMDH to help locate memory leak problems. A self-extracting executable is included that contains the following tools:

• Umdh.exe– This utility is used to dump the heap allocation information for a process.
• Umdh_nt4.exe and Dbghelp.dll – This is the Windows NT 4.0 SP6 version of the UMDH tool.
• Gflags.exe– This utility sets the appropriate registry entries for the application that will be analyzed. The operating system looks at the registry entries to determine if an application’s heap allocations will be tracked.
• Tlist.exe – This application lists all of the processes running on a machine and all of their related process IDs.
• Dhcmp.exe – This tool is used to compare two UMDH dumps to determine where a possible memory leak is occurring.
• Dhcmpgui.zip – This tool also compares two UMDH logs but has a user interface to make it easier to retrieve information. The .zip file contains the Dhcmpgui.exe file and its source code.

You can download this package either directly from the Microsoft Download Center or from our server:

Umdhtools.exe Microsoft Download Center Release Date: Feb. 5, 2001

Umdhtools.exe Download From Our Server Release Date: Feb. 5, 2001

Both we and Microsoft used the most current virus detection software available on the date of posting to scan this file for viruses. Once posted, the file is housed on secure servers that prevent any unauthorized changes to the file.

Before You Begin Using UMDH!

If you think you are experiencing a memory leak, please be aware that memory leaks may not be what they appear to be. You may discover that a memory leak is really not a memory leak at all, but rather a performance enhancement. For example, the Microsoft Jet database engine can consume large amounts of memory (up to 128 MB on a 256-MB computer) for its data retrieval and writing caches. The cache allows the Jet engine to achieve fast read-ahead and write-ahead buffering.

To determine whether or not a process is experiencing memory leaks, use the Windows Performance Monitor (Perfmon.exe) and monitor “Private Bytes” under the Process category for your application. Privates Bytes is the total amount of memory the process has allocated but is not sharing with other processes. Note that this is different from “Virtual Bytes,” which is also interesting to monitor. Virtual Bytes is the current size in bytes of the virtual address space the process is using. An application can leak virtual memory but not see a difference in the amount of private bytes allocated. If you don’t see memory rising when monitoring private bytes but suspect that you are still running out of memory, you should monitor virtual bytes to see if you are exhausting virtual memory.

The following MSDN article “Finding Leaks and Bottlenecks with a Windows NT PerfMon COM Object” will assist you further with detecting memory leaks:

Finding Leaks and Bottlenecks with Windows NT PerfMon COM Object

To be certain your application is leaking memory, you should place the suspect code in a loop with many iterations, and monitor private and virtual bytes for any increases of memory. Watch to make certain that the number of private bytes and virtual bytes doesn’t eventually stay the same and stops rising. If there is a point at which the memory stops rising, that is, it does not continue to climb indefinitely, you are not seeing a memory leak but are probably seeing some cache that is growing to its maximum size.

Capturing Heap Dumps with UMDH

UMDH (or Umdh_nt4.exe if you are using Windows NT 4.0 SP6) is a tool that dumps information about a process’s heap allocations. This information includes the callstack for each allocation, the number of allocations made through that callstack, and the number of bytes consumed through that callstack. For example:

  00005320 bytes in 0x14 allocations (@ 0x00000428) by: BackTrace00053
        ntdll!RtlDebugAllocateHeap+0x000000FD
        ntdll!RtlAllocateHeapSlowly+0x0000005A
        ntdll!RtlAllocateHeap+0x00000808
        MyApp!_heap_alloc_base+0x00000069
        MyApp!_heap_alloc_dbg+0x000001A2
        MyApp!_nh_malloc_dbg+0x00000023
        MyApp!_nh_malloc+0x00000016
        MyApp!operator new+0x0000000E
        MyApp!LeakyFunc+0x0000001E
        MyApp!main+0x0000002C
        MyApp!mainCRTStartup+0x000000FC
        KERNEL32!BaseProcessStart+0x0000003D 

This UMDH output shows that there were 21280 (0x5320) bytes allocated total from the callstack. The 21280 bytes were allocated from 20 (0x14) separate allocations of 1064 bytes (0x428). The callstack is given a identifier of BackTrace00053. To produce a dump of the heap allocations, you must first let the operating system know that you would like the kernel to track the allocations. This is done by using the Gflags.exe utility. Let’s assume that you want to dump the heap contents for Notepad.exe. By the way, most processes have more than one heap active and the UMDH log will contain all of them. First you need to enable stack trace acquisition for the application you want to test. This feature is not enabled by default. The command to enable it is:

gflags -i notepad.exe +ust

The command does not enable stack tracing for processes that are already running, but all future executions of Notepad.exe will have it enabled. Alternatively, you can set the flag through the GFLAGS user interface (run Gflags.exe without any arguments to get the user interface).

Before using UMDH, you also need to install the correct debug symbols for the components of your application as well as the operating system. Typically, the operating system symbols are installed in the SYMBOLS folder on the WINNT folder. UMDH attempts to find the symbol files by using the _NT_SYMBOL_PATH. The command for setting the path from a command prompt is:

set _NT_SYMBOL_PATH=%windir%\symbols

To learn more about setting up symbolic debugging, see the “Debug Symbols” section below.

With the image flags set and the symbols installed, you are ready to start Notepad. After the program is started, you need to determine the Process ID (PID) of the Notepad process just started. The command for this is:

tlist

You can find the PID from the output of the TLIST application. The PID information can also be obtained from Task Manager. Let’s assume the PID for the Notepad process you just started is 124. You can use UMDH to get a heap dump by using the following command:

umdh -p:124 -f:notepad124.log

Now you have a complete heap dump of the Notepad process in the Notepad124.log file.

Using Dhcmp.exe to Compare UMDH Logs While the UMDH log file contains valuable information about the current state of the heaps for a process, if you are concerned with finding a memory leak, it may be more valuable to compare the outputs of two logs and find out what callstack has seen the largest growth between the two dumps. The Dhcmp.exe utility helps compare two UMDH logs to provide an analysis of the difference between them. Once you have two logs captured at different intervals, you can then use the following command:

DHCMP dh1.log dh2.log > cmp12.txt

The .txt files will compare the effect on memory of running the suspect memory hog application. The output of the file generated by DHCMP resembles the following:

+    5320 (  f110 -  9df0)     3a allocs        BackTrace00053
Total increase == 5320 

For each BackTrace in the UMDH log files, there is a comparison made between the two logs files. This case shows that the last log file specified in the DHCMP command line had F110 bytes allocated while the first log in the DHCMP command line had 9DF0 bytes allocated for the same BackTrace (callstack). The 5320 is the difference in the number of bytes allocated. In this case, there were 5320 more bytes allocated between the time the two logs were captured. The bytes came from the callstack identified by BackTrace00053.

The next step is to find out what’s in that backtrace. If you open the second DH log file and search for “BackTrace00053” you might find something that resembles the following:

00005320 bytes in 0x14 allocations (@ 0x00000428) by: BackTrace00053
        ntdll!RtlDebugAllocateHeap+0x000000FD
        ntdll!RtlAllocateHeapSlowly+0x0000005A
        ntdll!RtlAllocateHeap+0x00000808
        MyApp!_heap_alloc_base+0x00000069
        MyApp!_heap_alloc_dbg+0x000001A2
        MyApp!_nh_malloc_dbg+0x00000023
        MyApp!_nh_malloc+0x00000016
        MyApp!operator new+0x0000000E
        MyApp!LeakyFunc+0x0000001E
        MyApp!main+0x0000002C
        MyApp!mainCRTStartup+0x000000FC
        KERNEL32!BaseProcessStart+0x0000003D 

By looking at the callstack, you can see that the LeakyFunc function is allocating memory by way of the Visual C++ run-time library. If you find that the number of allocations grows as you took more dumps, you might be able to conclude that memory is not being freed.

Enabling Stack Traces

The most important information in UMDH logs is the stack traces of the heap allocations. By analyzing them it can be understood whether a process leaks heap memory. These stack traces are not acquired by default. The feature can be enabled per-process or system-wide. Use the following command to enable stack tracing system-wide:

gflags -r +ust

Restart the computer after this command.

For per-process enabling, the command is

gflags -i APPNAME +ust

Where APPNAME is the file name of the executable including the extension (Services.exe, lsass.exe, and so on). The command will not enable stack tracing for a process that is already running. For this reason, in the case of processes that can’t be restarted (services, lsass, winlogon), the test compute must be restarted.

Use the following commands to verify what settings have been set system-wide or for a specific process:

System-wide:

gflags -r

Specific process:

gflags -i APP-NAME

By default, the maximum stack trace depth is 16. If you want to see deeper callstacks, you can increase this by running GFLAGS. Click to select System Registry and type a new depth in the Max. Stack Trace Capture Depth edit control. Click Apply, and then restart the computer.

IMPORTANT: If you are using Windows NT 4.0 Service Pack 6, you must use Umdh_nt4.exe, rather than Umdh.exe, and you must use the above command to set system-wide stack tracing. Be certain to reboot. Umdh_nt4 stack tracing does not work on a per process basis on Windows NT version 4. It must be set for the entire system.

Debug Symbols

UMDH uses the IMAGEHLP library for symbol manipulation. This is the standard tool for such tasks and it is also used by Microsoft debuggers. One of the most important steps to using UMDH is to ensure that you have good symbol files (.dbg or .pdb file) to get a good stack trace. The debug symbol files can be installed from the Windows NT CD, or in the case of Microsoft Windows 2000, from the Windows 2000 Support Tools CD. At a minimum, you need the Kernel32.dbg and Ntdll.dbg symbol files. You can always acquire additional debugging symbols as needed as you find out more about which components are leaking the memory.

When building components with Visual C++, it is important that you not have Program Database for Edit and Continue selected for the C++ compiler options. Instead, select Program Database.

To set the symbol path, initialize the _NT_SYMBOL_PATH environment variable to the path to be used. This can be done once when the test computer is set up by using the System application in Control Panel (Advanced, Environment Variables) or inside a command window before UMDH is run. Usually the symbols are copied to the test machine in the %windir%\Symbols folder. In this case, use the following command to set the variable:

set _NT_SYMBOL_PATH=%windir%\symbols

NOTE: Include the path to the PDBs for your application’s components as well.

Invoking the UMDH Tool

The only required command-line parameter for UMDH is the -p option, which specifies the PID of the process from which a heap dump will be obtained. The PID can be obtained by using Task Manager or the Tlist.exe program. For a command similar to the following, the log will be dumped to the standard output:

umdh -p:PID

UMDH also displays various informational messages to standard error, and therefore if not redirected it will be mixed with the real log. To gather the UMDH informational messages in a file, use the following command:

umdh -p:PID 2>umdh.msg

If you want to gather the log dumped by UMDH in a file, use one of the following commands:

umdh -p:PID >umdh.log

-or-

umdh -p:PID -f:umdh.log

The commands are completely equivalent.

The default log obtained by UMDH contains an enumeration of heap consumers sorted by allocation count. If, for debugging purposes, you also need a dump of all allocated blocks with their corresponding stack traces, the -d option must be used:

umdh -p:PID -d

If the log contains too much information, it can be limited only to big users that have the allocation count above a certain threshold. Use the following command:

umdh -p:PID -t:THRESHOLD

All the command-line options (-p, -f, -t, -d) can be specified simultaneously in any order. Following is a complicated command-line example:

umdh -p:123 -t:1000 -f:umdh.log -d

This command dumps the heaps for the process with PID 123 into the Umdh.log file. It dumps only stack traces that account for more than 1000 allocations and it also dumps the addresses of the heap blocks allocated through each stack trace.

UMDH Output Explained

If you redirected the log to a file (umdh -p:PID -f:umdh.log), the contents will resemble the following, which was obtained from a running Notepad process:

UMDH: Logtime 2000-06-28 10:54 - Machine=MYMachine - PID=704
*********** Heap 00270000 Information ********************
    Flags: 58000062
    Number Of Entries: 87
    Number Of Tags: 
    Bytes Allocated: 00008DF0
    Bytes Committed: 0000A000
    Total FreeSpace: 00001210
    Number of Virtual Address chunks used: 1
    Address Space Used: 
    Entry Overhead: 8
    Creator:  (Backtrace00007)
        ntdll!RtlDebugCreateHeap+0x00000196
        ntdll!RtlCreateHeap+0x0000023F
        ntdll!LdrpInitializeProcess+0x00000369
        ntdll!LdrpInitialize+0x0000028D
        ntdll!KiUserApcDispatcher+0x00000007
*********** Heap 00270000 Hogs ********************
000001A0 bytes in 0x4 allocations (@ 0x00000068) by: BackTrace00031
        ntdll!RtlDebugAllocateHeap+0x000000FB
        ntdll!RtlAllocateHeapSlowly+0x0000005B
        ntdll!RtlAllocateHeap+0x00000D81
        ntdll!LdrpAllocateDataTableEntry+0x00000039
        ntdll!LdrpMapDll+0x000002A4
        ntdll!LdrpLoadImportModule+0x0000010D
        ntdll!LdrpWalkImportDescriptor+0x0000008B
        ntdll!LdrpLoadImportModule+0x0000011D
        ntdll!LdrpWalkImportDescriptor+0x0000008B
        ntdll!LdrpLoadImportModule+0x0000011D
        ntdll!LdrpWalkImportDescriptor+0x0000008B
        ntdll!LdrpInitializeProcess+0x000009DC
        ntdll!LdrpInitialize+0x0000028D
        ntdll!KiUserApcDispatcher+0x00000007

000001A0 bytes in 0x4 allocations (@ 0x00000068) by: BackTrace00034
        ntdll!RtlDebugAllocateHeap+0x000000FB
        ntdll!RtlAllocateHeapSlowly+0x0000005B
        ntdll!RtlAllocateHeap+0x00000D81
        ntdll!LdrpAllocateDataTableEntry+0x00000039
        ntdll!LdrpMapDll+0x000002A4
        ntdll!LdrpLoadImportModule+0x0000010D
        ntdll!LdrpWalkImportDescriptor+0x0000008B
        ntdll!LdrpLoadImportModule+0x0000011D
        ntdll!LdrpWalkImportDescriptor+0x0000008B
        ntdll!LdrpLoadImportModule+0x0000011D
        ntdll!LdrpWalkImportDescriptor+0x0000008B
        ntdll!LdrpLoadImportModule+0x0000011D
        ntdll!LdrpWalkImportDescriptor+0x0000008B
        ntdll!LdrpInitializeProcess+0x000009DC
        ntdll!LdrpInitialize+0x0000028D
        ntdll!KiUserApcDispatcher+0x00000007 

The log contains a dump of every heap in the process. In this example, the log starts with a heap at address 270000. After a few global counters for the heap, the log contains a dump in decreasing sorted order of stack traces that are responsible for the most allocations. By comparing the dynamics of memory used at different moments you can deduce what happened in the process and if any heap use looks like a leak.

Problems That Can Be Encountered When Using UMDH The two most common errors when using UMDH are not having stack tracing enabled and not having good symbols. Incorrect symbols for Ntdll.dll prevent UMDH from running. For the other symbols files, UMDH will run but the log file will contain stack traces that do not have function names but instead have relative addresses inside modules. A distant third error is specifying a wrong PID.

The following error message results when you attempt to run UMDH for a process that does not have stack tracing enabled:

C:\>umdh -p:1140 UMDH: Logtime 2000-06-28 12:43 – Machine=MyMachine – PID=1140 Connecting…..Module enumeration complete. SymGetSymFromName(process, ntdll!RtlpStackTraceDataBase, xxx) failed, LastError = 126 UmdhGetAddrFromName couldn’t find Stack Trace DB pointer (ntdll!RtlpStackTraceDataBase).

ntdll.dll symbols are incorrect; we must be able to see non-import symbols.

Use the following command to double-check the settings for the process you are investigating:

gflags -i APPNAME

Use the following command in the case where you are relying on system-wide stack tracing:

gflags -r

These commands show the list of flags set for the application. Note that in the case of system-wide stack tracing, the feature might appear as active but if you did not restart the computer after running the gflags -r +ust command, it is not actually activated. If you want to know every application that has stack tracing enabled, you can look at the USTEnabled key under the following registry key:

HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows NT\CurrentVersion\Image File Execution Options

If you run UMDH on a process that has stack tracing enabled but have not restarted the application since setting the flags, you may get the following message in your logs:

A stack trace was not saved for this allocation (Index == 0)

If you run UMDH without setting the correct symbol path or the symbols are incorrect, you may get an error message in the log. However, you may just get incorrect or misleading callstacks. To verify whether you have the correct symbols, launch the NTSD system debugger for a process, for example:

ntsd notepad

Then, from the debugger console, run the LD command to load the symbol information for the module and the LM command to list what modules have their symbols loaded. If the output of the LM command shows export symbols loaded, the symbols are not good. If you have PDB symbols loaded, the symbols are good. You may get the following error message if you specified the wrong PID:

C:\>umdh -p:1000 UMDH: Logtime 2000-06-28 09:45 – Machine=MyMachine – PID=1000 Connecting… OpenProcess failed, LastError = 0x57

Additional Tools

Using DHCMPGUI

The Dhcmpgui.exe tool can be used to compare two UMDH logs. This tool is located in the Dhcmpgui.zip file. It calls out to Dhcmp.exe (make certain that you have this in your PATH). Click File 1 and File 2 to specify the two UMDH logs to compare.

After selecting the two files to compare, you will see a list of the differences. The columns listed are the number of bytes allocated, number of allocations, and the BackTraceID, respectively. You can sort by any one of those columns by clicking the appropriate button, and then clicking Sort. To see the BackTrace (callstack) for any line, just double-click it. This provides an easy way to compare two logs and to quickly get to the callstack for the callstack that is leaking memory.

The source code has been provided for Dhcmpgui.exe in case you want to make any modifications. The project requires Visual C++ version 6.0.

Calling UMDH from Visual Basic There may be times when it is useful to dump a number of logs over time because the leak may not be very noticeable at first. For example, suppose you suspect that your ASP application is leaking memory. It may be helpful to write a COM component in Visual Basic that shells out to UMDH. You can then call that component from your ASP page.

Following is some Visual Basic code that invokes UMDH and creates a log file based on the current time:

   Private Declare Function GetCurrentProcessId Lib "kernel32" () As Long
   Public Function GetProcessID()
   GetProcessID = GetCurrentProcessId()
   End Function  
.
.
.
   Dim strTime As String

   Dim sProcID As String
   sProcID = GetProcessID()
   strTime = "MYLOG_" & Format(Now(), "hhmm")
  
   Shell ("C:\UMDH\umdh -p:" & sProcID & " -f:d:\logs\" & strTime & ".txt") 

Last updated on February 25, 2001