AtomicFU

The idiomatic way to use atomic operations in Kotlin.

• Code it like AtomicReference/Int/Long, but run it in production efficiently as AtomicXxxFieldUpdater on Kotlin/JVM and as plain unboxed values on Kotlin/JS.
• Use Kotlin-specific extensions (e.g. inline updateAndGet and getAndUpdate functions).
• Compile-time dependency only (no runtime dependencies).
  • Post-compilation bytecode transformer that declares all the relevant field updaters for you on Kotlin/JVM.
  • Post-compilation JavaScript files transformer on Kotlin/JS.
• Multiplatform:
  • Kotlin/Native is supported.
  • However, Kotlin/Native works as library dependency at the moment (unlike Kotlin/JVM and Kotlin/JS).
  • This enables writing common Kotlin code with atomics that compiles for JVM, JS, and Native.
• Gradle for all platforms and Maven for JVM are supported.
• Additional features include:
  • JDK9 VarHandle.
  • Arrays of atomic values.
  • User-defined extensions on atomics
  • Locks
  • Testing of lock-free data structures.
  • Tracing operations

Example

Let us declare a top variable for a lock-free stack implementation:

import kotlinx.atomicfu.* // import top-level functions from kotlinx.atomicfu

private val top = atomic<Node?>(null) 

Use top.value to perform volatile reads and writes:

fun isEmpty() = top.value == null  // volatile read
fun clear() { top.value = null }   // volatile write

Use compareAndSet function directly:

if (top.compareAndSet(expect, update)) ... 

Use higher-level looping primitives (inline extensions), for example:

top.loop { cur ->   // while(true) loop that volatile-reads current value 
   ...
}

Use high-level update, updateAndGet, and getAndUpdate, when possible, for idiomatic lock-free code, for example:

fun push(v: Value) = top.update { cur -> Node(v, cur) }
fun pop(): Value? = top.getAndUpdate { cur -> cur?.next } ?.value

Declare atomic integers and longs using type inference:

val myInt = atomic(0)    // note: integer initial value
val myLong = atomic(0L)  // note: long initial value   

Integer and long atomics provide all the usual getAndIncrement, incrementAndGet, getAndAdd, addAndGet, and etc operations. They can be also atomically modified via += and -= operators.

Dos and Don'ts

• Declare atomic variables as private val or internal val. You can use just (public) val, but make sure they are not directly accessed outside of your Kotlin module (outside of the source set). Access to the atomic variable itself shall be encapsulated.
• Only simple operations on atomic variables directly are supported.
  • Do not read references on atomic variables into local variables, e.g. top.compareAndSet(...) is Ok, while val tmp = top; tmp... is not.
  • Do not leak references on atomic variables in other way (return, pass as params, etc).
• Do not introduce complex data flow in parameters to atomic variable operations, i.e. top.value = complex_expression and top.compareAndSet(cur, complex_expression) are not supported (more specifically, complex_expression should not have branches in its compiled representation). Extract complex_expression into a variable when needed.
• Use the following convention if you need to expose the value of atomic property to the public:

private val _foo = atomic<T>(initial) // private atomic, convention is to name it with leading underscore
public var foo: T by _foo            // public delegated property (val/var)

Gradle build setup

Building with Gradle is supported for all platforms.

JVM

You will need Gradle 4.10 or later. Add and apply AtomicFU plugin. It adds all the corresponding dependencies and transformations automatically. See additional configuration if that needs tweaking.

buildscript {
    ext.atomicfu_version = '0.17.2'

    dependencies {
        classpath "org.jetbrains.kotlinx:atomicfu-gradle-plugin:$atomicfu_version"
    }
}

apply plugin: 'kotlinx-atomicfu'

JS

Configure add apply plugin just like for JVM.

Native

This library is available for Kotlin/Native (atomicfu-native). Kotlin/Native uses Gradle metadata and needs Gradle version 5.3 or later. See Gradle Metadata 1.0 announcement for more details. Apply the corresponding plugin just like for JVM.

Atomic references for Kotlin/Native are based on FreezableAtomicReference and every reference that is stored to the previously frozen (shared with another thread) atomic is automatically frozen, too.

Since Kotlin/Native does not generally provide binary compatibility between versions, you should use the same version of Kotlin compiler as was used to build AtomicFU. See gradle.properties in AtomicFU project for its kotlin_version.

Common

If you write a common code that should get compiled or different platforms, add org.jetbrains.kotlinx:atomicfu to your common code dependencies or apply kotlinx-atomicfu plugin that adds this dependency automatically:

dependencies {
    compile "org.jetbrains.kotlinx:atomicfu:$atomicfu_version"
}

IR transformation for Kotlin/JS

There is a new option to turn on IR transformation for Kotlin/JS backend. You can add kotlinx.atomicfu.enableIrTransformation=true to your gradle.properties file in order to enable it.

Here is how transformation is performed for different JS compiler modes with this option enabled:

• kotlin.js.compiler=legacy: JavaScript transformer from the library is applied to the final compiled *.js files.
• kotlin.js.compiler=ir: compiler plugin transformations are appiled to the generated IR.
• kotlin.js.compiler=both: compiler plugin transformations are appiled to all compilations of IR targets, while compilations of legacy targets are transformed by the library.

Additional configuration

To set configuration options you should create atomicfu section in a build.gradle file, like this:

atomicfu {
  dependenciesVersion = '0.17.2'
}

JVM transformation options

To turn off transformation for Kotlin/JVM set option transformJvm to false.

Configuration option jvmVariant defines the Java class that replaces atomics during bytecode transformation. Here are the valid options:

• FU – atomics are replaced with AtomicXxxFieldUpdater.
• VH – atomics are replaced with VarHandle, this option is supported for JDK 9+.
• BOTH – multi-release jar file will be created with both AtomicXxxFieldUpdater for JDK <= 8 and VarHandle for JDK 9+.

JS transformation options

To turn off transformation for Kotlin/JS set option transformJs to false.

Here are all available configuration options (with their defaults):

atomicfu {
  dependenciesVersion = '0.17.2' // set to null to turn-off auto dependencies
  transformJvm = true // set to false to turn off JVM transformation
  jvmVariant = "FU" // JVM transformation variant: FU,VH, or BOTH 
  jsVariant = "JS" // JS transformation variant: JS or IR
  verbose = false // set to true to be more verbose  
}

Maven build setup

Declare AtomicFU version:

<properties>
     <atomicfu.version>0.17.2</atomicfu.version>
</properties> 

Declare provided dependency on the AtomicFU library (the users of the resulting artifact will not have a dependency on AtomicFU library):

<dependencies>
    <dependency>
        <groupId>org.jetbrains.kotlinx</groupId>
        <artifactId>atomicfu</artifactId>
        <version>${atomicfu.version}</version>
        <scope>provided</scope>
    </dependency>
</dependencies>

Configure build steps so that Kotlin compiler puts classes into a different classes-pre-atomicfu directory, which is then transformed to a regular classes directory to be used later by tests and delivery.

<build>
    <plugins>
        <!-- compile Kotlin files to staging directory -->
        <plugin>
            <groupId>org.jetbrains.kotlin</groupId>
            <artifactId>kotlin-maven-plugin</artifactId>
            <version>${kotlin.version}</version>
            <executions>
                <execution>
                    <id>compile</id>
                    <phase>compile</phase>
                    <goals>
                        <goal>compile</goal>
                    </goals>
                    <configuration>
                        <output>${project.build.directory}/classes-pre-atomicfu</output>
                    </configuration>
                </execution>
            </executions>
        </plugin>
        <!-- transform classes with AtomicFU plugin -->
        <plugin>
            <groupId>org.jetbrains.kotlinx</groupId>
            <artifactId>atomicfu-maven-plugin</artifactId>
            <version>${atomicfu.version}</version>
            <executions>
                <execution>
                    <goals>
                        <goal>transform</goal>
                    </goals>
                    <configuration>
                        <input>${project.build.directory}/classes-pre-atomicfu</input>
                        <!-- "VH" to use Java 9 VarHandle, "BOTH" to produce multi-version code -->
                        <variant>FU</variant>  
                    </configuration>
                </execution>
            </executions>
        </plugin>
    </plugins>
</build>

Additional features

AtomicFU provides some additional features that you can optionally use.

Arrays of atomic values

You can declare arrays of all supported atomic value types. By default arrays are transformed into the corresponding java.util.concurrent.atomic.Atomic*Array instances.

If you configure variant = "VH" an array will be transformed to plain array using VarHandle to support atomic operations.

val a = atomicArrayOfNulls<T>(size) // similar to Array constructor

val x = a[i].value // read value
a[i].value = x // set value
a[i].compareAndSet(expect, update) // do atomic operations

User-defined extensions on atomics

You can define you own extension functions on AtomicXxx types but they must be inline and they cannot be public and be used outside of the module they are defined in. For example:

@Suppress("NOTHING_TO_INLINE")
private inline fun AtomicBoolean.tryAcquire(): Boolean = compareAndSet(false, true)

Locks

This project includes kotlinx.atomicfu.locks package providing multiplatform locking primitives that require no additional runtime dependencies on Kotlin/JVM and Kotlin/JS with a library implementation for Kotlin/Native.

• SynchronizedObject is designed for inheritance. You write class MyClass : SynchronizedObject() and then use synchronized(instance) { ... } extension function similarly to the synchronized function from the standard library that is available for JVM. The SynchronizedObject superclass gets erased (transformed to Any) on JVM and JS, with synchronized leaving no trace in the code on JS and getting replaced with built-in monitors for locking on JVM.

• ReentrantLock is designed for delegation. You write val lock = reentrantLock() to construct its instance and use lock/tryLock/unlock functions or lock.withLock { ... } extension function similarly to the way jucl.ReentrantLock is used on JVM. On JVM it is a typealias to the later class, erased on JS.

Condition variables (notify/wait and signal/await) are not supported.

Testing lock-free data structures on JVM

You can optionally test lock-freedomness of lock-free data structures using LockFreedomTestEnvironment class. See example in LockFreeQueueLFTest. Testing is performed by pausing one (random) thread before or after a random state-update operation and making sure that all other threads can still make progress.

In order to make those test to actually perform lock-freedomness testing you need to configure an additional execution of tests with the original (non-transformed) classes for Maven:

<build>
    <plugins>
        <!-- additional test execution with surefire on non-transformed files -->
        <plugin>
            <artifactId>maven-surefire-plugin</artifactId>
            <executions>
                <execution>
                    <id>lockfree-test</id>
                    <phase>test</phase>
                    <goals>
                        <goal>test</goal>
                    </goals>
                    <configuration>
                        <classesDirectory>${project.build.directory}/classes-pre-atomicfu</classesDirectory>
                        <includes>
                            <include>**/*LFTest.*</include>
                        </includes>
                    </configuration>
                </execution>
            </executions>
        </plugin>
    </plugins>
</build>

For Gradle there is nothing else to add. Tests are always run using original (non-transformed) classes.

Tracing operations

You can debug your tests tracing atomic operations with a special trace object:

private val trace = Trace()
private val current = atomic(0, trace)

fun update(x: Int): Int {           
    // custom trace message
    trace { "calling update($x)" }
    // automatic tracing of modification operations 
    return current.getAndAdd(x)
}

All trace messages are stored in a cyclic array inside trace.

You can optionally set the size of trace's message array and format function. For example, you can add a current thread name to the traced messages:

private val trace = Trace(size = 64) {   
    index, // index of a trace message 
    text   // text passed when invoking trace { text }
    -> "$index: [${Thread.currentThread().name}] $text" 
} 

trace is only seen before transformation and completely erased after on Kotlin/JVM and Kotlin/JS.