- Windows Server 2008 R2
Svchost.exe, that black box amongst many other black boxes. If you ever happened to be in the business of watching what your scripts are getting up to on a Sunday morning and you are using Microsoft.XmlHttp, then you might be in for a surprise.
Every 2 hours a batch process on a group of servers fires off a script that in turn iteratively runs a second VBS script some 200-300 times. The script, calls a Web Service and performs a push/pull of instructions. Within a few days of the patch Tuesday reboot, you start noticing that memory use is going up, and up, and up.
You’ve done all of your deallocations, right? “set xmlHttp = nothing”? Yep, but despite that, memory use continues to grow. The culprit, svchost.exe. It grows until it’s into the page file and then grows a little bit more. Every run of the script puts between 4 and 100KB onto the memory footprint. At the end of the month, the servers are groaning because of memory starvation and your SAN array’s are not happy because of all of the paging.
I have been able to reproduce this on 3 separate and wholly independent Server 2008 systems (read different clients, enterprise/retail licensing, server hardware and install images) as well as on related servers (read from the same image on same or similar hardware). I have attempted to reproduce it on Windows Server 2012 R2 and I was not successful. Server 2012 R2 does not appear to be impacted by the issue. Running the iterator loop below for 10 minutes yields no increase in the memory use curve on the operating system, just a constant cycle of assign, release, assign, release that you would expect to see.
After a lot of diagnostics and a lot of me initially assuming that the problem was the web service (many, many wasted hours… although I did find a few bugs in the service code itself…) I managed to narrow it down to Microsoft.XmlHttp. More specifically, it’s in the way that CScript or WScript interfaces with Microsoft.XmlHttp at initialisation.
As you probably know, svchost itself is just a service wrapper. Inspection of the wrapper reveals a number of services running inside the wrapper. In this case the specific services are:
- COM+ Event System
- Windows Font Cache Service
- Network List Service
- Network Store Interface Service
- Secure Socket Tunneling Protocol Service
- WinHTTP Web Proxy Auto-Discovery
There are two things here that could be interesting, COM+ Event System and WinHTTP Web Proxy. Microsoft.XMLHTTP itself relies upon the WinHTTP stack for operation, but we are also using a COM interface to call it from VBScript.
While we cannot shutdown the COM+ Event Service and expect the operating system to survive for long, we can the WinHTTP Web Proxy Auto-Discovery Service. Did it release the memory consumed in the leak? No. So in the balance of probabilities, it’s coming from COM+.
The problem with that is in the need to reboot the server to safely clear the memory leak, hence why Patch Tuesday has been the true savior in keeping a gradual performance bottle neck from becoming a full scale meltdown. So what is going on?
I stripped off all of the web service and customisation parts and went back to vanilla Microsoft implementation examples. We cannot get much simpler than this.
Save it to a VBS and run it via CScript, run it a lot. Run it in a BAT file loop
Watch the svchost.exe processes until you spot the instance with the rising service working set (or private set). Now you know which one to focus on.
It’s memory leaked. Hold on, we’ve created the instance of Microsoft.XmlHttp (which is actually an instance of IXMLHTTPRequest), done something and told CScript to deallocate it (set xml = nothing). Why is it leaking memory?
The third parameter on .Open() is bAsync – is it an asynchronous request? It’s false above, meaning that the request is synchronous. It continues to leak. It would be more likely to leak asynchronously than synchronously, however changing that to true makes no difference.
So where is the leak being triggered? By process of line elimination we can reveal that the memory is committed into the svchost wrapper during xml.send(). Run it without .Send() as below and there is no growth in the scvhost process memory footprint no matter how many times you run it..
In the MSDN documentation for the .Send() method, it states
“If the input type is a BSTR, the response is always encoded as UTF-8. The caller must set a Content-Type header with the appropriate content type and include a charset parameter.”
So far we haven’t done that and we are sending a VBString – which is ultimately a BSTR in C++, so add in the necessary setRequestHeader beneath the .Open() method call in case it is a case of not following the documentation:
It isn’t. There is no change, it still results in an increase in process memory after cscript.exe has shutdown.
We have confirmed that there is a memory leak, where it is and what is triggering it. We can also be confident that given the extremely simple nature of the sample code printed above – and its match to the samples documentation – that it is being implemented correctly.
So the next step is to try and prove that there is an issue in the COM implementation between CreateObject and set nothing. This is achieved by running the allocate/deallocate (set/set nothing) in a loop as shown below
At this point you would expect to see a large increase in the svchost.exe memory footprint.
It does not happen.
1000 iterations and instantiation of the IXMLHTTPRequest later and there is no obvious exponential increase in the memory footprint of svchost.exe. It simply increments once i.e. the additional memory consumption is no worse than running the script with only 1 call to CreateObject/set nothing despite the fact that .send() has been called 1000 times.
What does that mean? Well, it would seem to suggest that the fault isn’t actually in IXMLHTTPRequest (Microsoft.XMLHTTP), but actually in VBScript itself. As a speculative suggestion, I would suggest that VBScript is registering event callbacks with COM+’s Event Management System on the first call to .Send() which are not being cleaned up by the garbage collector when “set nothing” is called in the code. So either there is a bug in VBScript or there is a bug in the event handling interface for COM+ event registration through which IXMLHTTPRequest is registering its own actions.
Most people aren’t going to notice this problem, they are morelikely to iterate instance of Microsoft.XmlHttp inside VBScript than they are to repeatedly externally iterate accross it. It just so happens that I need to fire it externally to the script processor via the command shell. The chances are that if you are reading this, so do you.
As of writing, I have not found a direct way to force VBScript to release the memory from scvhost, short of rebooting (or migrating to Windows Server 2012). Calling Microsoft.XmlHttp from WScript or CScript seems to be the problem and the fact that the web service scripts are using an external iterator to repeatedly call n new instances of CScript are exacerbating the situation. Simply put, the transaction load is the catalyst for spotting the leak. In most cases growth would be very subtle as would growth were the iteration internal to the CScript.exe script instance.
While not necessarily ideal, if you are in the position of being able to change provider, you can substitute Microsoft.XMLHTTP for MSXML2.ServerXmlHttp, which provides most of the functionality without making use if WinHTTP. This provider does not exhibit the memory growth issue as in its client counterpart, however its use requires MSXML 3 or 6 and you lose some functionality.
The fact that I could not reproduce the issue under Windows Server 2012 R2 suggests that the culprit has been fixed – either intentionally or inadvertantly. By default, Microsoft.XMLHTTP is a COM Class ID reference to msxml3.dll. Under Windows Server 2008 R2 the file version is SP11 at 8.110.7601.18334, under 2012 R2 the file version is simply 8.110.9600.16483. Yet oddly, with all systems fully patched, vbscript.dll under Windows Server 2008 R2 is version 5.8.9600.17041 (KB2929437) while its counterpart under Server 2012 R2 is 5.8.9600.17031.
What I can tell you is that these systems have been running this recursion script every 2 hours since the beginning of 2012 and the issue has only been observed in more recent months, therefore I suspect that Microsoft have a regression bug on their hands. Until it is fixed however, I have a load of (thankfully firewalled, private network) web service that have a DOS vulnerability. So do you.