Thread security & knowledge races

Earlier than we dive in to Swift actors, let’s have a simplified recap of laptop idea first.

An occasion of a pc program is known as course of). A course of incorporates smaller directions which are going to be executed in some unspecified time in the future in time. These instruction duties will be carried out one after one other in a serial order or concurrently. The working system is utilizing a number of threads) to execute duties in parallel, additionally schedules the order of execution with the assistance of a scheduler). 🕣

After a process is being accomplished on a given thread), the CPU can to maneuver ahead with the execution move. If the brand new process is related to a unique thread, the CPU has to carry out a context change. That is fairly an costly operation, as a result of the state of the previous thread must be saved, the brand new one must be restored earlier than we will carry out our precise process.

Throughout this context switching a bunch of different oprations can occur on totally different threads. Since trendy CPU architectures have a number of cores, they’ll deal with a number of threads on the similar time. Issues can occur if the identical useful resource is being modified on the similar time on a number of threads. Let me present you a fast instance that produces an unsafe output. 🙉

var unsafeNumber: Int = 0
DispatchQueue.concurrentPerform(iterations: 100) { i in
    print(Thread.present)
    unsafeNumber = i
}

print(unsafeNumber)

In case you run the code above a number of instances, it is potential to have a unique output every time. It’s because the concurrentPerform technique runs the block on totally different threads, some threads have greater priorities than others so the execution order is just not assured. You possibly can see this for your self, by printing the present thread in every block. A few of the quantity modifications occur on the principle thread, however others occur on a background thread. 🧵

The primary thread is a particular one, all of the consumer interface associated updates ought to occur on this one. If you’re making an attempt to replace a view from a background thread in an iOS software you may may get an warning / error or perhaps a crash. If you’re blocking the principle thread with an extended working software your complete UI can change into unresponsive, that is why it’s good to have a number of threads, so you possibly can transfer your computation-heavy operations into background threads.

It is a quite common strategy to work with a number of threads, however this may result in undesirable knowledge races, knowledge corruption or crashes resulting from reminiscence points. Sadly a lot of the Swift knowledge varieties aren’t thread protected by default, so if you wish to obtain thread-safety you normally needed to work with serial queues or locks to ensure the mutual exclusivity of a given variable.

var threads: [Int: String] = [:]
DispatchQueue.concurrentPerform(iterations: 100) { i in
    threads[i] = "(Thread.present)"
}
print(threads)

The snippet above will crash for certain, since we’re making an attempt to switch the identical dictionary from a number of threads. That is known as a data-race. You possibly can detect these form of points by enabling the Thread Sanitizer below the Scheme > Run > Diagnostics tab in Xcode. 🔨

Now that we all know what’s a knowledge race, let’s repair that by utilizing a daily Grand Central Dispatch primarily based strategy. We’ll create a brand new serial dispatch queue to forestall concurrent writes, this may syncronize all of the write operations, however in fact it has a hidden price of switching the context every time we replace the dictionary.

var threads: [Int: String] = [:]
let lockQueue = DispatchQueue(label: "my.serial.lock.queue")
DispatchQueue.concurrentPerform(iterations: 100) { i in
    lockQueue.sync {
        threads[i] = "(Thread.present)"
    }
}
print(threads)

This synchronization method is a fairly fashionable answer, we may create a generic class that hides the inner personal storage and the lock queue, so we will have a pleasant public interface that you need to use safely with out coping with the inner safety mechanism. For the sake of simplicity we’re not going to introduce generics this time, however I’ll present you a easy AtomicStorage implementation that makes use of a serial queue as a lock system. 🔒

import Basis
import Dispatch

class AtomicStorage {

    personal let lockQueue = DispatchQueue(label: "my.serial.lock.queue")
    personal var storage: [Int: String]
    
    init() {
        self.storage = [:]
    }
        
    func get(_ key: Int) -> String? {
        lockQueue.sync {
            storage[key]
        }
    }
    
    func set(_ key: Int, worth: String) {
        lockQueue.sync {
            storage[key] = worth
        }
    }

    var allValues: [Int: String] {
        lockQueue.sync {
            storage
        }
    }
}

let storage = AtomicStorage()
DispatchQueue.concurrentPerform(iterations: 100) { i in
    storage.set(i, worth: "(Thread.present)")
}
print(storage.allValues)

Since each learn and write operations are sync, this code will be fairly sluggish for the reason that complete queue has to attend for each the learn and write operations. Let’s repair this actual fast by altering the serial queue to a concurrent one, and marking the write perform with a barrier flag. This fashion customers can learn a lot sooner (concurrently), however writes might be nonetheless synchronized via these barrier factors.

import Basis
import Dispatch

class AtomicStorage {

    personal let lockQueue = DispatchQueue(label: "my.concurrent.lock.queue", attributes: .concurrent)
    personal var storage: [Int: String]
    
    init() {
        self.storage = [:]
    }
        
    func get(_ key: Int) -> String? {
        lockQueue.sync {
            storage[key]
        }
    }
    
    func set(_ key: Int, worth: String) {
        lockQueue.async(flags: .barrier) { [unowned self] in
            storage[key] = worth
        }
    }

    var allValues: [Int: String] {
        lockQueue.sync {
            storage
        }
    }
}

let storage = AtomicStorage()
DispatchQueue.concurrentPerform(iterations: 100) { i in
    storage.set(i, worth: "(Thread.present)")
}
print(storage.allValues)

In fact we may pace up the mechanism with dispatch limitations, alternatively we may use an os_unfair_lock, NSLock or a dispatch semaphore to create related thread-safe atomic objects.

One necessary takeaway is that even when we are attempting to pick out the most effective out there possibility by utilizing sync we’ll at all times block the calling thread too. Which means that nothing else can run on the thread that calls synchronized features from this class till the inner closure completes. Since we’re synchronously ready for the thread to return we won’t make the most of the CPU for different work. ⏳

We are able to say that there are various issues with this strategy:

• Context switches are costly operations
• Spawning a number of threads can result in thread explosions
• You possibly can (by accident) block threads and stop additional code execution
• You possibly can create a impasse if a number of duties are ready for one another
• Coping with (completion) blocks and reminiscence references are error inclined
• It is very easy to overlook to name the right synchronization block

That is various code simply to offer thread-safe atomic entry to a property. Even though we’re utilizing a concurrent queue with limitations (locks have issues too), the CPU wants to change context each time we’re calling these features from a unique thread. As a result of synchronous nature we’re blocking threads, so this code is just not probably the most environment friendly.

Happily Swift 5.5 provides a protected, trendy and general a lot better various. 🥳

Introducing Swift actors

Now let’s refactor this code utilizing the new Actor kind launched in Swift 5.5. Actors can defend inside state via knowledge isolation guaranteeing that solely a single thread could have entry to the underlying knowledge construction at a given time. Lengthy story quick, every part inside an actor might be thread-safe by default. First I am going to present you the code, then we’ll discuss it. 😅

import Basis

actor AtomicStorage {

    personal var storage: [Int: String]
    
    init() {
        self.storage = [:]
    }
        
    func get(_ key: Int) -> String? {
        storage[key]
    }
    
    func set(_ key: Int, worth: String) {
        storage[key] = worth
    }

    var allValues: [Int: String] {
        storage
    }
}

Activity {
    let storage = AtomicStorage()
    await withTaskGroup(of: Void.self) { group in
        for i in 0..<100 {
            group.async {
                await storage.set(i, worth: "(Thread.present)")
            }
        }
    }
    print(await storage.allValues)
}

To start with, actors are reference varieties, similar to lessons. They will have strategies, properties, they’ll implement protocols, however they do not assist inheritance.

Since actors are intently associated to the newly launched async/await concurrency APIs in Swift you ought to be acquainted with that idea too if you wish to perceive how they work.

The very first huge distinction is that we need not present a lock mechanism anymore in an effort to present learn or write entry to our personal storage property. Which means that we will safely entry actor properties throughout the actor utilizing a synchronous method. Members are remoted by default, so there’s a assure (by the compiler) that we will solely entry them utilizing the identical context.

What is going on on with the brand new Activity API and all of the await key phrases? 🤔

Properly, the Dispatch.concurrentPerform name is a part of a parallelism API and Swift 5.5 launched concurrency as a substitute of parallelism, now we have to maneuver away from common queues and use structured concurrency to carry out duties in parallel. Additionally the concurrentPerform perform is just not an asynchronous operation, it will block the caller thread till all of the work is completed throughout the block.

Working with async/await implies that the CPU can work on a unique process when awaits for a given operation. Each await name is a possible suspension level, the place the perform may give up the thread and the CPU can carry out different duties till the awaited perform resumes & returns with the required worth. The new Swift concurrency APIs are constructed on prime a cooperative thread pool, the place every CPU core has simply the correct quantity of threads and the suspension & continuation occurs “just about” with the assistance of the language runtime. That is much more environment friendly than precise context switching, and in addition implies that while you work together with async features and await for a perform the CPU can work on different duties as a substitute of blocking the thread on the decision facet.

So again to the instance code, since actors have to guard their inside states, they solely permits us to entry members asynchronously while you reference from async features or outdoors the actor. That is similar to the case once we had to make use of the lockQueue.sync to guard our learn / write features, however as a substitute of giving the flexibility to the system to carry out different duties on the thread, we have totally blocked it with the sync name. Now with await we may give up the thread and permit others to carry out operations utilizing it and when the time comes the perform can resume.

Inside the duty group we will carry out our duties asynchronously, however since we’re accessing the actor perform (from an async context / outdoors the actor) now we have to make use of the await key phrase earlier than the set name, even when the perform is just not marked with the async key phrase.

The system is aware of that we’re referencing the actor’s property utilizing a unique context and now we have to carry out this operation at all times remoted to eradicate knowledge races. By changing the perform to an async name we give the system an opportunity to carry out the operation on the actor’s executor. Afterward we’ll have the ability to outline customized executors for our actors, however this function is just not out there but.

At present there’s a international executor implementation (related to every actor) that enqueues the duties and runs them one-by-one, if a process is just not working (no rivalry) it will be scheduled for execution (primarily based on the precedence) in any other case (if the duty is already working / below rivalry) the system will simply pick-up the message with out blocking.

The humorous factor is that this doesn’t needed implies that the very same thread… 😅

import Basis

extension Thread {
    var quantity: String {
        "(worth(forKeyPath: "personal.seqNum")!)"
    }
}

actor AtomicStorage {

    personal var storage: [Int: String]
    
    init() {
        print("init actor thread: (Thread.present.quantity)")
        self.storage = [:]
    }
        
    func get(_ key: Int) -> String? {
        storage[key]
    }
    
    func set(_ key: Int, worth: String) {
        storage[key] = worth + ", actor thread: (Thread.present.quantity)"
    }

    var allValues: [Int: String] {
        print("allValues actor thread: (Thread.present.quantity)")
        return storage
    }
}


Activity {
    let storage = AtomicStorage()
    await withTaskGroup(of: Void.self) { group in
        for i in 0..<100 {
            group.async {
                await storage.set(i, worth: "caller thread: (Thread.present.quantity)")
            }
        }
    }    
    for (okay, v) in await storage.allValues {
        print(okay, v)
    }
}

Multi-threading is difficult, anyway similar factor applies to the storage.allValues assertion. Since we’re accessing this member from outdoors the actor, now we have to await till the “synchronization occurs”, however with the await key phrase we may give up the present thread, wait till the actor returns again the underlying storage object utilizing the related thread, and voilá we will proceed simply the place we left off work. In fact you possibly can create async features inside actors, while you name these strategies you may at all times have to make use of await, irrespective of if you’re calling them from the actor or outdoors.

There’s nonetheless lots to cowl, however I do not wish to bloat this text with extra superior particulars. I do know I am simply scratching the floor and we may discuss non-isolated features, actor reentrancy, international actors and lots of extra. I am going to positively create extra articles about actors in Swift and canopy these matters within the close to future, I promise. Swift 5.5 goes to be an incredible launch. 👍

Hopefully this tutorial will assist you to begin working with actors in Swift. I am nonetheless studying lots in regards to the new concurrency APIs and nothing is written in stone but, the core workforce remains to be altering names and APIs, there are some proposals on the Swift evolution dashboard that also must be reviewed, however I feel the Swift workforce did an incredible job. Thanks everybody. 🙏

Truthfully actors appears like magic and I already love them. 😍