Reading 23: Mutual Exclusion

Recall that race conditions arise when multiple concurrent computations share access (reading and writing) to the same mutable data at the same time. This is unsafe, because the correctness of the program may depend on accidents of timing of the concurrent operations.

There are three general ways to solve the problem:

• Don’t share. Don’t share mutable data between concurrent computations. Instead, each computation might have its own copy. This strategy is sometimes called confinement.
• Don’t mutate. Never write to the data, only read it; in other words, use immutability.
• Don’t access at the same time. Ensure that the critical parts of the concurrent computations, which read and write the shared mutable data, can’t possibly interleave. This strategy is called mutual exclusion or synchronization.

This reading discusses all three ways, but focuses mostly on the last one, mutual exclusion. We will look at these techniques in several contexts of concurrent computation:

• Asynchronous functions in TypeScript, using objects in memory
• Workers in TypeScript, using files on disk

As a running example, let’s use a simple bank account type:

/** Represents a mutable bank account, with an integer balance in dollars. */
interface BankAccount { 

    /** deposit $1 into the account */
    deposit():void;

    /** withdraw $1 from the account */
    withdraw():void;

    /** account balance, an integer number of dollars, possibly negative */
    readonly balance: number;
}

We will define different implementations of the BankAccount type that may store the account balance in different ways, depending on the context.

As a simulation of concurrent computations accessing the same mutable data, we’ll go back to our cash machine example. A single global bank account will be shared by multiple cash machines. Each cash machine repeatedly deposits a dollar and withdraws a dollar. Here is the (non-concurrent) version of the code:

const account:BankAccount = new BankAccount(200);

// each cash machine runs a bunch of transactions on the single shared account
const NUMBER_OF_TRANSACTIONS = 10;
function cashMachine() {
    console.log('started with balance', account.balance)
    for (let i = 0; i < NUMBER_OF_TRANSACTIONS; ++i) {
        account.deposit(); // put a dollar in
        account.withdraw(); // take a dollar out
    }
    console.log('finished with balance', account.balance);
}

// run a bunch of cash machines
const NUMBER_OF_CASH_MACHINES = 5;
for (let i = 0; i < NUMBER_OF_CASH_MACHINES; ++i) {
    cashMachine();
}

Since a cash machine withdraws each dollar right after depositing it, we should expect at the end of the day that there was no effect on the account balance – and indeed that’s the case when we run this simple straight-line code, since it has no concurrency at the moment. Each call to cashMachine() runs to completion before the next call to cashMachine() starts:

started with balance 200
finished with balance 200
started with balance 200
finished with balance 200
started with balance 200
finished with balance 200
started with balance 200
finished with balance 200
started with balance 200
finished with balance 200

Let’s start by looking at a fortunately rosy case: asynchronous functions.

Suppose we change cashMachine() into an async function:

async function cashMachine():Promise<void> {
    console.log('started with balance', account.balance)
    for (let i = 0; i < NUMBER_OF_TRANSACTIONS; ++i) {
        account.deposit(); // put a dollar in
        account.withdraw(); // take a dollar out
    }
    console.log('finished with balance', account.balance);
}

Here’s what we see when we run it:

started with balance 200
finished with balance 200
started with balance 200
finished with balance 200
started with balance 200
finished with balance 200
started with balance 200
finished with balance 200
started with balance 200
finished with balance 200

Nothing actually changed! The code still behaves like the synchronous cashMachine() function did. This is because the body of cashMachine() doesn’t contain any await expression, and await is the only place that an asynchronous function can give up control in the middle of its execution. So each cashMachine() still runs to completion, and returns, before the next cashMachine() is even called.

Let’s make the asynchronous function more interesting. Suppose the cashMachine() function is simulating a person at a cash machine, doing these transactions. We’ll put in some random delays to simulate human variability:

async function cashMachine():Promise<void> {
    console.log('started with balance', account.balance)
    for (let i = 0; i < NUMBER_OF_TRANSACTIONS; ++i) {
        await timeout(Math.random()*100);
        account.deposit(); // put a dollar in
        await timeout(Math.random()*100);
        account.withdraw(); // take a dollar out
    }
    console.log('finished with balance', account.balance);
}

Now we see some more interesting execution traces, for example:

started with balance 200
started with balance 200
started with balance 200
started with balance 200
started with balance 200
finished with balance 203
finished with balance 203
finished with balance 201
finished with balance 201
finished with balance 200

Some of the balances in this trace are not 200, which means those balances were printed while one or more of the cash machines were in the middle of their deposit/withdrawal action sequence. That’s not necessarily an error. We should expect the account balance to fluctuate while these transactions are being made. The essential correctness question is: at the end of the day, after all the cash machines are done, do we still have $200 in the account? In this trace, that’s true: the last line shows $200 at the end.

So the asynchronous calls to cashMachine() are now interleaving with each other, so the key question becomes: is there a race condition? Is there a particular interleaving that might leave the account balance at some value different than $200? Or that might throw an exception?

The answer is no, the code is safe from bad interleavings, and the reason is that an async function can only lose control at an await. JavaScript runs asynchronous functions using only a single thread of control, which has to switch back and forth among the active concurrent computations. For async functions, those switches happen only when the function reaches an await. When the function is in between awaits, it has sole control, and it cannot be interrupted or interleaved with another asynchronous function.

For the bank account example, that means each cash machine’s critical deposit() and withdraw() operations never run at the same time as another cash machine’s:

await timeout(Math.random()*100);  // <=== might switch to another cash machine here
account.deposit();                 // <=== runs without interruption
await timeout(Math.random()*100);  // <=== might switch to another cash machine here
account.withdraw();                // <=== runs without interruption

… so we never find ourselves in the situation that caused bad interleavings.

Now let’s look at the bank-account example in the context of Worker, JavaScript’s lightweight process abstraction.

We’ll start each cash machine as a fresh worker:

// run a bunch of cash machines as Workers
const NUMBER_OF_CASH_MACHINES = 5;
for (let i = 0; i < NUMBER_OF_CASH_MACHINES; ++i) {
    new Worker('./cash-machine-worker.js');
}

where cash-machine-worker.js just does this:

cashMachine();

What happens when we run this? Here’s an example trace:

started with balance 200
finished with balance 200
started with balance 200
finished with balance 200
started with balance 200
started with balance 200
started with balance 200
finished with balance 200
finished with balance 200
finished with balance 200

There is clearly interleaving happening here; the Workers are indeed running concurrently, and starting and finishing independently. But the balances are suspicious. Why are they always 200? Why don’t we ever see fluctuating balances, like we saw with the asynchronous function using random delays?

To shed light, let’s change our cash machine transactions so that they only deposit, and never withdraw:

account.deposit(); // put a dollar in
// account.withdraw(); // but don't take a dollar out, this time

We should now expect the account to end the day with $200 plus the total number of $1 deposits that were done, which is the number of cash machines (5) times the number of transactions per machine (10). And that’s what the synchronous (non-concurrent) version of the code now does:

started with balance 200
finished with balance 210
started with balance 210
finished with balance 220
started with balance 220
finished with balance 230
started with balance 230
finished with balance 240
started with balance 240
finished with balance 250

But the worker version does something different:

started with balance 200
finished with balance 210
started with balance 200
finished with balance 210
started with balance 200
started with balance 200
started with balance 200
finished with balance 210
finished with balance 210
finished with balance 210

Why?

The new Workers we created do not share the bank account object. Mutable objects are confined within the worker where they were created. So there’s no risk of a race condition – but we also aren’t doing what we wanted, which is to have a single bank account that all the cash machines are transacting.

Workers are more suited to message-passing concurrency, so message-passing would be a more appropriate way for them to use the BankAccount type, as we will see in the next reading. But to continue to explore the issues with shared-memory concurrency, let’s change our bank account so that it uses mutable data that is shared by workers, and other processes as well: the filesystem.

Here is an implementation of BankAccount that keeps the account balance in a text file, so that different workers can share access to the file. The type has internal operations for reading and writing the file:

private writeBalance(balance:number):void {
    fs.writeFileSync(this.filename, balance.toString());
}

private readBalance():number {
    const fileData = fs.readFileSync(this.filename).toString();
    return parseInt(fileData);
}

And then the deposit, withdraw, and balance operations use these read and write methods:

public deposit():void {
    let balance:number = this.readBalance();
    balance = balance + 1;
    this.writeBalance(balance);
}

public withdraw():void {
    let balance:number = this.readBalance();
    balance = balance - 1;
    this.writeBalance(balance);
}

public get balance():number {
    return this.readBalance();
}

Using the synchronous (non-concurrent) cash machine code that makes deposits only, this file-backed account works as expected, increasing the account by a total of $50:

started with balance 200
finished with balance 210
started with balance 210
finished with balance 220
started with balance 220
finished with balance 230
started with balance 230
finished with balance 240
started with balance 240
finished with balance 250

But the worker version of the code does not work as hoped; it usually finishes with something less than $250 in the account:

started with balance 200
started with balance 200
started with balance 201
started with balance 203
started with balance 201
finished with balance 203
finished with balance 210
finished with balance 216
finished with balance 218
finished with balance 219

The workers are interleaving, money is disappearing from the account, and we get a different answer every time we run it. Now we have a race condition.

It turns out we can experience this race condition in the asynchronous-function version of our code as well, if deposit and withdraw are asynchronous methods.

The file-manipulation code above used readFileSync and writeFileSync, which are synchronous functions. But in the asynchronous-function model, we should write this code using the asynchronous versions of these functions, which return promises:

private async writeBalance(balance:number):Promise<void> {
    return fs.writeFile(this.filename, balance.toString());
}

private async readBalance():Promise<number> {
    const fileData = ( await fs.readFile(this.filename) ).toString();
    return parseInt(fileData);
}

That means the methods of the bank account also must become asynchronous:

public async deposit():Promise<void> {
    let balance:number = await this.readBalance();
    balance = balance + 1;
    await this.writeBalance(balance);
}

public async withdraw():Promise<void> {
    let balance:number = await this.readBalance();
    balance = balance - 1;
    await this.writeBalance(balance);
}

public get balance():Promise<number> {
    return this.readBalance();
}

Now we have the potential for a race condition, because the critical mutator methods deposit and withdraw contain await expressions where they might lose control, and interleave with other deposits or withdrawals. And indeed, if we run the code to deposit a total of $50 from 5 different asynchronous cashMachine() functions, we see the same kind of bad behavior that the worker-version was seeing, for example:

started with balance 200
started with balance 200
started with balance 200
started with balance 200
started with balance 200
finished with balance 212
finished with balance 212
finished with balance 212
finished with balance 212
finished with balance 213

The asynchronous filesystem bank account suffers from a race condition.

The correctness of a concurrent program should not depend on accidents of timing.

Since race conditions caused by concurrent manipulation of shared mutable data are disastrous bugs — hard to discover, hard to reproduce, hard to debug — we need a way for concurrent modules that share mutable data to synchronize with each other.

Locks are one synchronization technique. A lock is an abstraction that allows at most one concurrent computation to own it at a time. Often the owner is a thread, but it might instead be a process, or worker, or asynchronous function. Holding a lock is how the thread informs other threads: “I’m working with this thing, don’t touch it right now.”

Locks have two operations:

• acquire allows a thread to take ownership of a lock. If a thread tries to acquire a lock currently owned by another thread, it blocks until the other thread releases the lock. At that point, it will contend with any other threads that are trying to acquire the lock. At most one thread can own the lock at a time.

• release relinquishes ownership of the lock, allowing another thread to take ownership of it.

Blocking means, in general, that a thread waits (without doing further work) until an event occurs. The await operator blocks waiting for a promise to fulfill. Synchronous I/O operations like readFileSync() and writeFileSync() block until the file operation is done.

An acquire(l) on thread 1 will block if another thread (say thread 2) is holding lock l. The event it waits for is thread 2 performing release(l). At that point, if thread 1 can acquire l, it continues running its code, with ownership of the lock. It is possible that another thread (say thread 3) was also blocked on acquire(l). If so, either thread 1 or 3 (the winner is nondeterministic) will take the lock l and continue. The other will continue to block, waiting for release(l) again.

Bank account example

When used properly and carefully, locks can prevent race conditions. But then another problem rears its ugly head. Because the use of locks requires threads to wait (acquire blocks when another thread is holding the lock), it’s possible to get into a situation where two threads are waiting for each other — and hence neither can make progress.

In the figure to the right, suppose A and B are making simultaneous transfers between two accounts in our bank.

A transfer between accounts needs to lock both accounts, so that money can’t disappear from the system. A and B each acquire the lock on their respective “from” account first: A acquires the lock on account 1, and B acquires the lock on account 2. Now, each must acquire the lock on their “to” account: so A is waiting for B to release the account 2 lock, and B is waiting for A to release the account 1 lock. Stalemate! A and B are frozen in a “deadly embrace,” and accounts are locked up.

Deadlock occurs when concurrent modules are stuck waiting for each other to do something. A deadlock may involve more than two modules, e.g., A may be waiting for B, which is waiting for C, which is waiting for A. None of them can make progress. The essential feature of deadlock is a cycle of dependencies like this.

You can also have deadlock without using any locks. For example, a message-passing system can experience deadlock when message buffers fill up. If a client fills up the server’s buffer with requests, and then blocks waiting to add another request, the server may then fill up the client’s buffer with results and then block itself. So the client is waiting for the server, and the server waiting for the client, and neither can make progress until the other one does. Again, deadlock ensues.

A locking discipline is a strategy for ensuring that code using locks is safe for concurrency. We must satisfy two conditions:

1. Every piece of shared mutable data must be guarded by some lock. The data may not be read or written unless the code doing the reading or writing is currently holding the lock.

2. If an invariant involves multiple pieces of shared mutable data (which might even be in different objects), then all the data involved must be guarded by the same lock. Once the lock has been acquired in order to mutate the data, the invariant must be reestablished before releasing the lock.

This approach is called the monitor pattern. A monitor is a class whose methods are mutually exclusive, so that only one thread can be inside an instance of the class at a time. Each instance of the class has a lock, and every public method acquires the lock when it starts, and releases the lock when it returns.

The monitor pattern satisfies both rules shown here. All the shared mutable data in the rep — which the rep invariant depends on — are guarded by the same lock.

Let’s look at how our filesystem bank account can use a lock to protect itself from concurrent access. Locks are not built into the JavaScript library or runtime system, so let’s use a third-party package, await-lock. This package provides a type AwaitLock with the usual two operations, acquire and release. The acquire operation happens to be called acquireAsync() to remind programmers that it is an asynchronous method, returning a promise that must be awaited before the lock is fully acquired.

To use this lock type in the monitor pattern, the bank account class first needs to create a lock for itself, and store it as a private instance variable:

private lock: AwaitLock = new AwaitLock();

Then every public method of the bank account acquires and releases the lock:

public async deposit():Promise<void> {
    await this.lock.acquireAsync();
    try {
        let balance:number = await this.readBalance();
        balance = balance + 1;
        await this.writeBalance(balance);
    } finally {
        this.lock.release();
    }
}

public async withdraw():Promise<void> {
    await this.lock.acquireAsync();
    try {
        let balance:number = await this.readBalance();
        balance = balance - 1;
        await this.writeBalance(balance);
    } finally {
        this.lock.release();
    }
}

Goals

Strategies

Producing a concurrent program that is safe from bugs, easy to understand, and ready for change requires careful thinking. Heisenbugs will skitter away as soon as you try to pin them down, so debugging simply isn’t an effective way to achieve correct code. And threads can interleave their operations in so many different ways that you will never be able to test even a small fraction of all possible executions.

• Acquiring a lock allows a thread to have exclusive access to the data guarded by that lock, forcing other threads to block — as long as those threads are also trying to acquire that same lock.

• The monitor pattern guards the rep of a datatype with a single lock that is acquired by every method.

• Blocking caused by acquiring multiple locks creates the possibility of deadlock.