Mono Migration Analyzer, Results

by Miguel de Icaza

Within 12 hours of announcing Moma we received 114 submissions generated by the tool detailing what features were missing.

There were a number of interesting discoveries, these are:

False Positives: One of the most commonly reported error were method calls and property accesses to the System.Net.WebRequest class. These turned out to be incorrect, since this class is an abstract class that happens to have stubbed all those methods and throws NotImplementedExceptions as a mechanism to ensure that implementors get those methods right.

Developers will always be using a subclass that is created by the WebRequest.Create factory method (It makes me wonder why the methods were not flagged as abstract)

New or overwritten methods: In a number of places, overwritten methods are showing up in the report. For example, a very common one is Exception.GetType(), or with Label::set_Text (the Label.Text setter), in both cases the problem is caused by 2.0 introducing a new method that overrides the base definition.

The code, if compiled with Mono would have produced the correct result, but the binaries would not have worked as the member references were done to newer classes. Most of these are innocuous, and we already introduced the missing APIs for the most commonly reported problems.

Outdated MonoTODOs: We have plenty of miss-used MonoTODOs, these are the messages that we used internally to flag that something in a method is incomplete, and ideally, it should provide an explanation of why the flag was set.

But this is an attribute that we have used a bit recklessly over the past few years, and has had a number of meanings, from valid uses like "This method only implements half the specification" to less useful instances like "I do not like this API", to "Someone should optimize this routine" to "One day I would like to look improve this". So plenty of MonoTODOs were hints aimed at the source code maintainer, and not really to the developer that consumes the API.

Eliminating all of the bogus MonoTODOs is an ongoing process, but we might still get a few erroneous reports for a while that might confuse developers.

In particular, a lot of people hit on various methods in DataSet. None of those errors are correct, they are implemented, but they had old "MonoTODO" flags left behind.

Limitations in the .NET API: In a number of cases, developers have resorted to P/Invoke into the Win32 API, and the results are very interesting, the top reasons for using P/Invoke have been:

  • Message dispatching: SendMessage, PostMessage.
  • GDI/DC operations: GetDC, CreateBrush, CreateRgn, SelectObject, DrawText.
  • Window operations: SetWindowPos, SetForegroundWindow ShowWindow, GetWindowRect, ClientToScreen, MoveWindow and Hook handling.
  • Allocation: GlobalAlloc
  • Fonts and Printing Dialogs.
  • Shell access: to get recently used files, icons for files and a handful of others.

There are of course many more, but these are the majority of the uses.

Some of those APIs are easily translated into portable calls (Message dispatching, window operations, allocation) while others are very hard (GDI/DC operations) and some others are better if replaced as a chunk with alternative code.

Our next goal will be to provide a portability assembly that will take care of those operations that are easily translated, so P/Invoke calls would be replaced with calls into this portability assembly. On Windows, we will continue to P/Invoke the same methods, while on other platforms we would route the request through Mono's internals.

In the case of Shell access, the best thing to do is to provide a cross-platform API that does the integration with the user shell, and provide an API that developers can migrate to. This would be effectively a "complement" for APIs that are today not available for Windows.Forms and even be a lot simpler for many people to use than resorting to P/Invoking.

Lack of Comments: The first version of Moma lacked support for entering some comments from the person submitting the report, so all we got in our hands are the list of unimplemented methods, but we do not know much about the applications that are using it.

When we ran this over applications we had, we were able to identify pieces that can be replaced as units based on the class names where they were used (this is part of the report that people get, but not part of the information transmitted to us). For example, a commercial application that we want to port (with the assistance of the owners) has neatly isolated things like printing dialogs in its own file.

Cecil and Obfuscated Assemblies: Some of the applications that were processed were obfuscated. But the P/Invoke signatures also seem obfuscated, am not sure how those are actually resolved at runtime.

Moma's version: at some point, we will need to enable Moma to download new API definitions from our site, to allow developers to check back if their applications are ready without having to download a new copy of Moma.

Accounting: There are a number of limitations on the accounting and statistics that can be done with the results.

It is hard to tell with precision how many of these are .NET 1.1 vs .NET 2.0, which ones are C# and VB and which ones are ASP.NET vs Windows.Forms. This is caused because Moma only uploads lists of missing methods, it does not upload other information that might be confidential, so Moma erred on the side of safety and removed data that was not necessary for us.

The accounting for ASP.NET is further limited, since Moma does work on assemblies, and not on .aspx files. Those that used ASP.NET are most likely helper code that is used in ASP.NET applications.

Another challenge with the data is that each submission contains the report for all the assemblies submitted. The end user might have chosen a single program, a single project (made up of many executables and libraries) or multiple programs.

This is not a problem, but it is worth keeping in mind when you read the results below.

Results

From the submitted applications, ten of them work out of the box without any changes; Another 30 contained reports that could be ignored (MonoTODOs that should not have been there, the NotImplementeds that did not exist). Almost all of them contain calls to methods that are either easy to implement, or were already implemented.

At least 29 of the applications are 2.0-based (by looking at Label::set_Text and Exception::GetType) and 16 of them are VisualBasic applications. One of the applications embeds IKVM.

32 applications were ASP.NET applications; 56 used Windows.Forms; 32 use System.Data and 44 used System.Drawing (see the above caveats about interpreting this data; This is based on assemblies and submissions, not on projects).

P/Invoke usage, from the 114 applications submitted:

  • 67 do not contain any P/Invoke calls.
  • 9 contain between 1 and 3 P/Invoke calls.
  • 10 contain between 4 and 10 P/Invoke calls.
  • 9 contain between 11 and 20 P/Invoke calls.
  • 5 contain between 22 and 48 P/Invoke calls.
  • 6 contain between 55 and 100 P/Invoke calls.
  • 5 contain between 102 and 180 P/Invoke calls.
  • 3 container more than 200 P/Invoke calls
    In the three cases, about half of the P/Invokes are to their own libraries.

(details).

So roughly half of the .NET applications can port without any concerns about P/Invoke. The remaining 40% are easily portable (those that have less than 48 P/Invokes) another 5% would take a few weeks, and maybe a Linux/Mono expert to assist on the port. The last 5% will require some large refactoring to work on Linux.

The application that uses the most P/Invoke that we have received so far seems to be some kind of designer and also happens to use the System.Security.CodeAccessPermission::Assert a lot.

Update: Sebastien comments that this is most likely a case of badly designed software, it is intended to avoid lots of stack walks to check for permissions. The above was probably copy/pasted from some recipe as an optimization that was useful in the 1.x frameworks. It is not required for 2.0.

From the received applications, 15 of them need more work on VisualBasic. Although most of the Visual Basic applications reported NotImplementedExceptions for a number of their methods in CompilerServices, that turned out to be false positives, which leaves us with only 24 methods missing in the VB.NET runtime

Unsupported Technologies

There are a few features that we do not support, and do not plan to support, in Mono. These technologies are either being phased out (EnterpriseServices, System.Messaging, to be replaced by WCF/Olive), or they would require a lot of work that we are currently not planning on doing (COM).

EnterpriseServicesis was used only by two applications, and all they seem to miss is a call to ServicedComponent::Dispose(). I suspect these are generated by a tool, considering the lack of any other methods references (and that we know are flagged).

System.Messaging is another of the libraries that we have not implemented and it showed up in only three applications.

COM, showed up in three applications.

Update

By the time this post went up this morning, we had received 171 submissions.

Posted on 28 Nov 2006