Bypassing Kotlin’s Null-Safety

In this short article, we will have a look at how we can bypass Kotlin’s native null-safety with sun.misc.Unsafe, and see why it can be dangerous even if we are not messing up with it directly.

Mythical sun.misc.Unsafe

The sun.misc.Unsafe class is an internal JVM tool for executing low-level operations like off-heap memory allocation, thread parking, CAS, and much more.

This class is like one of those scary computer game creatures that are there only to intimidate us, in theory, we can’t get close because they are part of the environment, but it’s often possible by exploiting glitches or holes.

If we try to access the Unsafe instance, we encounter a private constructor and a static getUnsafe() method that raises a SecurityException practically every time we call it:

public final class Unsafe {
    private static final Unsafe theUnsafe;
    // ...

    private Unsafe() {}

    @CallerSensitive
    public static Unsafe getUnsafe() {
        Class var0 = Reflection.getCallerClass();
        if (!VM.isSystemDomainLoader(var0.getClassLoader())) {
            throw new SecurityException("Unsafe");
        } else {
            return theUnsafe;
        }
    }
}

 

So, in theory, it’s guarded by a strong encapsulation, and an exception being thrown on every getUnsafe() call… but we do have the Reflection mechanism, and we can easily bypass those:

private fun getUnsafe(): Unsafe {
    return Unsafe::class.java.getDeclaredField("theUnsafe")
            .apply { isAccessible = true }
            .let { it.get(null) as Unsafe }
}

Mighty Unsafe.allocateInstance()

This method allocates an empty instance of a given class directly on the heap ignoring field initialization and constructors.

And this allows us, indeed, to effectively bypass Kotlin’s safety checks:

A cool thing to do on Friday’s evening, but what about just not using Unsafe and staying (null)safe?

Problem: Unsafe in External Libraries

The problem is that most Java libraries were written with Java in mind, where using Unsafe for certain scenarios is slightly less unsafe than it is e.g., for Kotlin.

This is especially the case with serialization/deserialization libraries – one of such is Google’s Gson which internally uses Unsafe for instantiating objects in certain situations – which is totally acceptable for Java.

If we start using it in Kotlin, we indeed might end up with an undesired behaviour observed above:

@Test
fun unsafe_2() {
    val foo = Gson().fromJson("{}", Foo::class.java)

    assertThat(foo.nonNullable).isNull()
}

In this case, we simply need to perform checks manually after instantiation, which is not super problematic – what’s problematic is the lack of consciousness that this happens, which can cost much.

Are you sure the library you are using is not doing that internally?

Code snippets can be found on GitHub.

Key Takeaways

  • Kotlin’s null-safety does not go beyond objects’ initialization phase and is bypassable
  • External libraries that use Unsafe internally can do that too – it’s important to be aware of this
You May Also Like

Spock, Java and Maven

Few months ago I've came across Groovy - powerful language for JVM platform which combines the power of Java with abilities typical for scripting languages (dynamic typing, metaprogramming).

Together with Groovy I've discovered spock framework (https://code.google.com/p/spock/) - specification framework for Groovy (of course you can test Java classes too!). But spock is not only test/specification framework - it also contains powerful mocking tools.

Even though spock is dedicated for Groovy there is no problem with using it for Java classes tests. In this post I'm going to describe how to configure Maven project to build and run spock specifications together with traditional JUnit tests.


Firstly, we need to prepare pom.xml and add necessary dependencies and plugins.

Two obligatory libraries are:
<dependency>
<groupid>org.spockframework</groupId>
<artifactid>spock-core</artifactId>
<version>0.7-groovy-2.0</version>
<scope>test</scope>
</dependency>
<dependency>
<groupid>org.codehaus.groovy</groupId>
<artifactid>groovy-all</artifactId>
<version>${groovy.version}</version>
<scope>test</scope>
</dependency>
Where groovy.version is property defined in pom.xml for more convenient use and easy version change, just like this:
<properties>
<gmaven-plugin.version>1.4</gmaven-plugin.version>
<groovy.version>2.1.5</groovy.version>
</properties>

I've added property for gmaven-plugin version for the same reason ;)

Besides these two dependencies, we can use few additional ones providing extra functionality:
  • cglib - for class mocking
  • objenesis - enables mocking classes without default constructor
To add them to the project put these lines in <dependencies> section of pom.xml:
<dependency>
<groupid>cglib</groupId>
<artifactid>cglib-nodep</artifactId>
<version>3.0</version>
<scope>test</scope>
</dependency>
<dependency>
<groupid>org.objenesis</groupId>
<artifactid>objenesis</artifactId>
<version>1.3</version>
<scope>test</scope>
</dependency>

And that's all for dependencies section. Now we will focus on plugins necessary to compile Groovy classes. We need to add gmaven-plugin with gmaven-runtime-2.0 dependency in plugins section:
<plugin>
<groupid>org.codehaus.gmaven</groupId>
<artifactid>gmaven-plugin</artifactId>
<version>${gmaven-plugin.version}</version>
<configuration>
<providerselection>2.0</providerSelection>
</configuration>
<executions>
<execution>
<goals>
<goal>compile</goal>
<goal>testCompile</goal>
</goals>
</execution>
</executions>
<dependencies>
<dependency>
<groupid>org.codehaus.gmaven.runtime</groupId>
<artifactid>gmaven-runtime-2.0</artifactId>
<version>${gmaven-plugin.version}</version>
<exclusions>
<exclusion>
<groupid>org.codehaus.groovy</groupId>
<artifactid>groovy-all</artifactId>
</exclusion>
</exclusions>
</dependency>
<dependency>
<groupid>org.codehaus.groovy</groupId>
<artifactid>groovy-all</artifactId>
<version>${groovy.version}</version>
</dependency>
</dependencies>
</plugin>

With these configuration we can use spock and write our first specifications. But there is one issue: default settings for maven-surefire plugin demand that test classes must end with "..Test" postfix, which is ok when we want to use such naming scheme for our spock tests. But if we want to name them like CommentSpec.groovy or whatever with "..Spec" ending (what in my opinion is much more readable) we need to make little change in surefire plugin configuration:
<plugin>
<groupid>org.apache.maven.plugins</groupId>
<artifactid>maven-surefire-plugin</artifactId>
<version>2.15</version>
<configuration>
<includes>
<include>**/*Test.java</include>
<include>**/*Spec.java</include>
</includes>
</configuration>
</plugin>

As you can see there is a little trick ;) We add include directive for standard Java JUnit test ending with "..Test" postfix, but there is also an entry for spock test ending with "..Spec". And there is a trick: we must write "**/*Spec.java", not "**/*Spec.groovy", otherwise Maven will not run spock tests (which is strange and I've spent some time to figure out why Maven can't run my specs).

Little update: instead of "*.java" postfix for both types of tests we can write "*.class" what is in my opinion more readable and clean:
<include>**/*Test.class</include>
<include>**/*Spec.class</include>
(thanks to Tomek Pęksa for pointing this out!)

With such configuration, we can write either traditional JUnit test and put them in src/test/java directory or groovy spock specifications and place them in src/test/groovy. And both will work together just fine :) In one of my next posts I'll write something about using spock and its mocking abilities in practice, so stay in tune.

[:en] Operational problems with Zookeeper

This post is a summary of what has been presented by Kathleen Ting on StrangeLoop conference. You can watch the original here: http://www.infoq.com/presentations/Misconfiguration-ZooKeeper I've decided to put this selection here for quick reference. ...This post is a summary of what has been presented by Kathleen Ting on StrangeLoop conference. You can watch the original here: http://www.infoq.com/presentations/Misconfiguration-ZooKeeper I've decided to put this selection here for quick reference. ...

Grails render as JSON catch

One of a reasons your controller doesn't render a proper response in JSON format might be wrong package name that you use. It is easy to overlook. Import are on top of a file, you look at your code and everything seems to be fine. Except response is still not in JSON format.

Consider this simple controller:

class RestJsonCatchController {
def grailsJson() {
render([first: 'foo', second: 5] as grails.converters.JSON)
}

def netSfJson() {
render([first: 'foo', second: 5] as net.sf.json.JSON)
}
}

And now, with finger crossed... We have a winner!

$ curl localhost:8080/example/restJsonCatch/grailsJson
{"first":"foo","second":5}
$ curl localhost:8080/example/restJsonCatch/netSfJson
{first=foo, second=5}

As you can see only grails.converters.JSON converts your response to JSON format. There is no such converter for net.sf.json.JSON, so Grails has no converter to apply and it renders Map normally.

Conclusion: always carefully look at your imports if you're working with JSON in Grails!

Edit: Burt suggested that this is a bug. I've submitted JIRA issue here: GRAILS-9622 render as class that is not a codec should throw exception