github.com/emaloney/CleanroomConcurrency

Utilities for simplifying asynchronous code execution & coordinating concurrent access to shared resources


License
MIT

Documentation

Gilt Tech logo

CleanroomConcurrency

CleanroomConcurrency provides utilities for simplifying asynchronous code execution and coordinating concurrent access to shared resources.

CleanroomConcurrency is part of the Cleanroom Project from Gilt Tech.

Swift 2.2 compatibility

The master branch of this project is Swift 2.2 compliant and therefore requires Xcode 7.3 or higher to compile.

License

CleanroomConcurrency is distributed under the MIT license.

CleanroomConcurrency is provided for your use—free-of-charge—on an as-is basis. We make no guarantees, promises or apologies. Caveat developer.

Adding CleanroomConcurrency to your project

Carthage compatible

You’ll need to integrate CleanroomConcurrency into your project in order to use the API it provides. You can choose:

Once integrated, just add the following import statement to any Swift file where you want to use CleanroomConcurrency:

import CleanroomConcurrency

Using CleanroomConcurrency

CleanroomConcurrency provides:

  • A small set of global asynchronous dispatching functions that declare a simplified interface for executing code asynchronously.

  • CriticalSection — A simple way to synchronize the execution of code across multiple threads.

  • ReadWriteCoordinator — A mechanism to coordinate access to mutable resources shared across multiple threads.

  • ThreadLocalValue — Simplifies access to thread-local values stored in the threadDictionary of the calling NSThread.

AsyncFunctions

The AsyncFunctions.swift file contains top-level functions that declare a simplified interface for executing code asynchronously.

Under the hood, these functions all rely on a single, privately-maintained concurrent Grand Central Dispatch (GCD) queue.

async

The async function provides a simple notation for specifying that a block of code should be executed asynchronously.

async takes as a parameter a no-argument function returning Void, and is typically invoked with a closure:

async {
    println("This will execute asynchronously")
}

The operation specified by the closure—the println() call above—will be executed asynchronously.

async with delay

A variation on the async function takes a delay parameter, an NSTimeInterval value specifying the minimum number of seconds to wait before asynchronously executing the closure:

async(delay: 0.35) {
    println("This will execute asynchronously after at least 0.35 seconds")
}

Note that this function does not perform real-time scheduling, so the asynchronous operation is not guaranteed to execute immediately once delay number of seconds has elapsed; instead, it will execute after at least delay number of seconds has elapsed.

mainThread

The mainThread function enqueues an operation for eventual execution on the main NSThread of the application.

This is useful because certain operations are only allowed to be executed on the main thread, such as view hierarchy manipulations. This is why the main thread is sometimes called the user interface thread.

The mainThread function is typically invoked with a closure as follows:

mainThread {
    UIApplication.sharedApplication().keyWindow!.rootViewController = self
}

The notation above ensures that the rootViewController is changed only on the main thread.

mainThread with delay

As with the async function, a variation of the mainThread function takes a delay parameter:

mainThread(delay: 0.35) {
    view.hidden = true
    view.alpha = 1.0
}

In the example above, the closure will be executed on the main thread after at least 0.35 seconds have elapsed.

asyncBarrier

The asyncBarrier function submits a barrier operation for asynchronous execution.

Barriers make it possible to create a synchronization point for operations:

  • Operations submitted prior to the submission of the barrier operation are guaranteed to execute before the barrier.

  • After all operations submitted prior to the barrier have been executed, then the barrier operation is executed.

  • While the barrier operation is executing, no other operations in the queue will execute.

  • Once the barrier operation finishes executing, normal concurrent behavior resumes; operations submitted after the barrier will then be executed.

Barrier operations are submitted for eventual execution using the notation:

asyncBarrier {
    println("When this executes, no other operations will be executing")
}

Important: Because these functions all rely on a single shared GCD queue, use of the asyncBarrier function can have application-wide impact. As a result, the function should be used sparingly, and only in situations where an application-wide barrier is truly needed.

If you do not need an application-wide barrier, it may be best to maintain your own queue and use the dispatch_barrier_async GCD function instead.

CriticalSection

A CriticalSection provides a simple way to synchronize the execution of code across multiple threads.

CriticalSections are a form of a mutex (or mutual exclusion) lock: when one thread is executing code within a given critical section, it is guaranteed that no other thread will be executing within the same instance.

In this way, CriticalSections are similar to @synchronized blocks in Objective-C.

Using a CriticalSection

With a CriticalSection, any code inside the execute closure (shown below as the "// code to execute" comment) is executed only after exclusive access to the critical section has been acquired by the calling thread.

let cs = CriticalSection()

cs.execute {
    // code to execute
}

Because it is possible for a CriticalSection to be used in a way where execute could block forever—resulting in a thread dealock—a variation is provided that allows a timeout to be specified:

let cs = CriticalSection()

let success = cs.executeWithTimeout(1.0) {
    // code to execute
}

if !success {
    // handle the fact that `executeWithTimeout()` timed out
}

The timeout is an NSTimeInterval value specifying the number of seconds to wait for access to the critical section before giving up.

In the example above, the calling thread may block for up to 1.0 seconds waiting for the critical section represented by cs to be acquired.

If cs can be acquired within that time, "// code to execute" will execute and executeWithTimeout() will return true.

If exclusive access to the critical section cs can't be acquired within 1.0 seconds, nothing will be executed and executeWithTimeout() will return false.

Note: It is best to design your implementation to avoid the potential for a deadlock. However, sometimes this is not possible, which is why the executeWithTimeout() function is provided.

Implementation Details

The CriticalSection implementation uses an NSRecursiveLock internally, which enables CriticalSections to be re-entrant. This means that a thread can't deadlock on a CriticalSection it already holds.

In addition, the CriticalSection implementation also performs internal exception trapping to ensure that the lock state remains consistent.

ReadWriteCoordinator

ReadWriteCoordinator instances can be used to coordinate access to a mutable resource shared across multiple threads.

You can think of the ReadWriteCoordinator as a dual read/write lock having the following properties:

  • The read lock allows any number of readers to execute concurrently.

  • The write lock allows one and only one writer to execute at a time.

  • As long as there is at least one reader executing, the write lock cannot be acquired.

  • As long as the write lock is held, no readers can execute.

  • All reads execute synchronously; that is, they block the calling thread until they complete.

  • All writes execute asynchronously.

  • Any read submitted before a write is guaranteed to be executed before that write, while any write submitted before a read is guaranteed to be executed before that read. This ensures a consistent view of shared resource's state.

The term lock is used in this document for conceptual clarity. In reality, the implementation uses Grand Central Dispatch and not a traditional lock.

Usage

For any given shared resource that needs to be protected by a read/write lock, you can create a ReadWriteCoordinator instance to manage access to that resource.

let lock = ReadWriteCoordinator()

You would then hold a reference to that ReadWriteCoordinator for the lifetime of the shared resource.

Reading

Whenever you need read-only access to the shared resource, you wrap your access within a call to the ReadWriteCoordinator's read() function, which is typically called with a trailing closure:

lock.read {
    // read some data from the shared resource
}

Because reads are executed synchronously, the ReadWriteCoordinator can be used within property getters, eg.:

var globalCount: Int {
    get {
        var count: Int?
        lock.read {
            count = // get a count from the shared resource
        }
        return count!
    }
}

Writing

Whenever you need to modify the state of the shared resource, you do so using the enqueueWrite() function of the ReadWriteCoordinator.

As with read(), this function is typically invoked with a closure:

lock.enqueueWrite {
    // modify the shared resource
}

Unlike read(), however, the enqueueWrite() function is asynchronous, as its name implies.

Write operations are submitted to the underlying GCD queue, and enqueueWrite() returns immediately.

When a write operation is enqueued, any already-pending read operations will be allowed to finish. Once the write lock can be acquired, the function passed to enqueueWrite() will be executed.

Under the hood, writes are submitted as asynchronous barrier operations to the receiver's GCD queue, ensuring that reads are always consistent with the order of writes.

ThreadLocalValue

ThreadLocalValue provides a mechanism for accessing thread-local values stored in the threadDictionary of the calling NSThread.

This implementation provides three main advantages over using the threadDictionary directly:

  • Type-safetyThreadLocalValue is implemented as a Swift generic, allowing it to enforce type safety.

  • Namespacing to avoid key clashes — To prevent clashes between different code modules using thread-local storage, ThreadLocalValues can be instantiated with a namespace used to construct the underlying threadDictionary key.

  • Use thread-local storage as a lockless cacheThreadLocalValues can be constructed with an optional instantiator that is used to construct values when the underlying threadDictionary doesn't have a value for the given key.

Namespacing

Namespacing can prevent key clashes when multiple subsystems need to share thread-local storage.

For example, two different subsystems may wish to store an NSDateFormatter instance in thread-local storage. If they were each to store their NSDateFormatters using the key "dateFormatter", for example, there would be a clash. The first value set for the "dateFormatter" key would always be overwritten by the second value set for that key.

Constructing your ThreadLocalValue with a namespace can prevent that:

let loggerDateFormatter = ThreadLocalValue<NSDateFormatter>(namespace: "Logger", key: "dateFormatter")

let saleDateFormatter = ThreadLocalValue<NSDateFormatter>(namespace: "SaleViewModel", key: "dateFormatter")

When a namespace is used, the ThreadLocalValue implementation constructs a value for its fullKey property by concatenating the values passed to the namespace and key parameters in the format "namespace.key".

In the example above, the loggerDateFormatter uses the fullKey "Logger.dateFormatter" while the saleDateFormatter uses "SaleViewModel.dateFormatter".

Because only the fullKey is used when accessing the underlying threadDictionary, these two ThreadLocalValue instances can each be used independently without their underlying values conflicting.

Regardless of whether a ThreadLocalValue uses a namespace, the key used to access the threadDictionary is always available via the fullKey property.

Thread-local caching

ThreadLocalValue instances can also be used to treat thread-local storage as a lockless cache.

Objects that are expensive to create, such as NSDateFormatter instances, can be cached in thread-local storage without incurring the locking overhead that would be required by an object cache shared among multiple threads.

This capability is available to ThreadLocalValues created with an instantiator function, and it works as follows:

  • If the value() function is called when the underlying threadDictionary doesn't have a value associated with the receiver's fullKey, the instantiator will be invoked to create a value.

  • If the instantiator returns a non-nil value, this value will be stored in the threadDictionary of the calling thread using the key fullKey of the ThreadLocalValue instance.

  • Future calls to the ThreadLocalValue's value() or cachedValue() functions will return the value created by the instantiator until the underlying value is changed.

Using a ThreadLocalValue instance to cache an NSDateFormatter would look like:

let df = ThreadLocalValue<NSDateFormatter>(namespace: "Events", key: "dateFormatter") { _ in
    let fmt = NSDateFormatter()
    fmt.locale = NSLocale(localeIdentifier: "en_US")
    fmt.timeZone = NSTimeZone(forSecondsFromGMT: 0)
    fmt.dateFormat = "yyyyMMdd_HHmmss"
    return fmt
}

In the example above, df is constructed with an instantiator closure. If df.value() is called when there is no NSDateFormatter associated with the key "Events.dateFormatter" in the calling thread's theadDictionary, the instantiator will be invoked to create a new NSDateFormatter.

Using thread-local storage as a cheap cache is best suited for cases where the long-term expense of acquiring read locks every time the object is accessed is greater than the expense of creating a new instance multiplied by the number of unique threads that will access the value.

API documentation

For detailed information on using CleanroomConcurrency, API documentation is available.

About

The Cleanroom Project began as an experiment to re-imagine Gilt’s iOS codebase in a legacy-free, Swift-based incarnation.

Since then, we’ve expanded the Cleanroom Project to include multi-platform support. Much of our codebase now supports tvOS in addition to iOS, and our lower-level code is usable on Mac OS X and watchOS as well.

Cleanroom Project code serves as the foundation of Gilt on TV, our tvOS app featured by Apple during the launch of the new Apple TV. And as time goes on, we'll be replacing more and more of our existing Objective-C codebase with Cleanroom implementations.

In the meantime, we’ll be tracking the latest releases of Swift & Xcode, and open-sourcing major portions of our codebase along the way.

Contributing

CleanroomConcurrency is in active development, and we welcome your contributions.

If you’d like to contribute to this or any other Cleanroom Project repo, please read the contribution guidelines.

Acknowledgements

API documentation for CleanroomConcurrency is generated using Realm’s jazzy project, maintained by JP Simard and Samuel E. Giddins.