Reading 21: Locks and Synchronization

We defined thread safety for a data type or a function as behaving correctly when used from multiple threads, regardless of how those threads are executed, without additional coordination.

Here’s the general principle: the correctness of a concurrent program should not depend on accidents of timing.

To achieve that correctness, we enumerated four strategies for making code safe for concurrency:

1. Confinement: don’t share data between threads, by keeping variables and the data they point to accessible from only one thread.
2. Immutability: make the shared data immutable, by using final variables and immutable types.
3. Use existing threadsafe data types: use a data type that does the coordination for you.
4. Synchronization: prevent threads from accessing the shared data at the same time. This is what we use to implement a threadsafe type, but we didn’t discuss it at the time.

In this reading, we’ll talk about strategy 4, using synchronization to implement your own data type that is safe for shared-memory concurrency.

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 memory to synchronize with each other.

Locks are one synchronization technique. A lock is an abstraction that allows at most one thread to own it at a time. Holding a lock is how one thread tells 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.

Using a lock also tells the compiler and processor that you’re using shared memory concurrently, so that registers and caches will be flushed out to shared storage. This avoids the problem of reordering, ensuring that the owner of a lock is always looking at up-to-date data.

Blocking means, in general, that a thread waits (without doing further work) until an event occurs.

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). Either thread 1 or 3 will take the lock l and continue. The other will continue to block, waiting for release(l) again.

Bank account example

Our first example of shared memory concurrency was a bank with cash machines. The diagram from that example is on the right.

The bank has several cash machines, all of which can read and write the same account objects in memory.

Of course, without any coordination between concurrent reads and writes to the account balances, things went horribly wrong.

To solve this problem using locks, we can add a lock that protects each bank account. Now, before they can access or update an account balance, cash machines must first acquire the lock on that account.

In the diagram to the right, both A and B are trying to access account 1. Suppose B acquires the lock first. Then A must wait to read or write the balance until B finishes and releases the lock. This ensures that A and B are synchronized, but another cash machine C is able to run independently on a different account (because that account is protected by a different lock).

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: 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: the signal feature of deadlock is a cycle of dependencies, e.g. A is waiting for B which is waiting for C which is waiting for A. None of them can make progress.

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.

Let’s see how to use synchronization to implement a threadsafe ADT.

You can see all the code for this example on GitHub: edit buffer example. You are not expected to read and understand all the code. All the relevant parts are excerpted below.

Suppose we’re building a multi-user editor, like Google Docs, that allows multiple people to connect to it and edit it at the same time. We’ll need a mutable datatype to represent the text in the document. Here’s the interface; basically it represents a string with insert and delete operations:

/** An EditBuffer represents a threadsafe mutable
 *  string of characters in a text editor. */
public interface EditBuffer {
    /**
     * Modifies this by inserting a string.
     * @param pos position to insert at
                      (requires 0 <= pos <= current buffer length)
     * @param ins string to insert
     */
    public void insert(int pos, String ins);

    /**
     * Modifies this by deleting a substring
     * @param pos starting position of substring to delete 
     *                (requires 0 <= pos <= current buffer length)
     * @param len length of substring to delete 
     *                (requires 0 <= len <= current buffer length - pos)
     */
    public void delete(int pos, int len);

    /**
     * @return length of text sequence in this edit buffer
     */
    public int length();

    /**
     * @return content of this edit buffer
     */
    public String toString();
}

A very simple rep for this datatype would just be a string:

public class SimpleBuffer implements EditBuffer {
    private String text;
    // Rep invariant:
    //   true
    // Abstraction function: 
    //   represents the sequence text[0],...,text[text.length()-1]

The downside of this rep is that every time we do an insert or delete, we have to copy the entire string into a new string. That gets expensive. Another rep we could use would be a character array, with space at the end. That’s fine if the user is just typing new text at the end of the document (we don’t have to copy anything), but if the user is typing at the beginning of the document, then we’re copying the entire document with every keystroke.

A more interesting rep, which is used by many text editors in practice, is called a gap buffer. It’s basically a character array with extra space in it, but instead of having all the extra space at the end, the extra space is a gap that can appear anywhere in the buffer. Whenever an insert or delete operation needs to be done, the datatype first moves the gap to the location of the operation, and then does the insert or delete. If the gap is already there, then nothing needs to be copied — an insert just consumes part of the gap, and a delete just enlarges the gap! Gap buffers are particularly well-suited to representing a string that is being edited by a user with a cursor, since inserts and deletes tend to be focused around the cursor, so the gap rarely moves.

/** GapBuffer is a non-threadsafe EditBuffer that is optimized
 *  for editing with a cursor, which tends to make a sequence of
 *  inserts and deletes at the same place in the buffer. */
public class GapBuffer implements EditBuffer {
    private char[] a;
    private int gapStart;
    private int gapLength;
    // Rep invariant:
    //   0 <= gapStart <= a.length
    //   0 <= gapLength <= a.length - gapStart
    // Abstraction function: 
    //   represents the sequence a[0],...,a[gapStart-1],
    //                           a[gapStart+gapLength],...,a[a.length-1]

In a multiuser scenario, we’d want multiple gaps, one for each user’s cursor, but we’ll use a single gap for now.

Steps to develop the datatype

Locks are so commonly-used that Java provides them as a built-in language feature.

In Java, every object has a lock implicitly associated with it — a String, an array, an ArrayList, and every class you create, all of their object instances have a lock. Even a humble Object has a lock, so bare Objects are often used for explicit locking:

Object lock = new Object();

You can’t call acquire and release on Java’s intrinsic locks, however. Instead you use the synchronized statement to acquire the lock for the duration of a statement block:

synchronized (lock) { // thread blocks here until lock is free
    // now this thread has the lock
    balance = balance + 1;
    // exiting the block releases the lock
}

Synchronized regions like this provide mutual exclusion: only one thread at a time can be in a synchronized region guarded by a given object’s lock. In other words, you are back in sequential programming world, with only one thread running at a time, at least with respect to other synchronized regions that refer to the same object.

Locks guard access to data

When you are writing methods of a class, the most convenient lock is the object instance itself, i.e. this. As a simple approach, we can guard the entire rep of a class by wrapping all accesses to the rep inside synchronized (this).

/** SimpleBuffer is a threadsafe EditBuffer with a simple rep. */
public class SimpleBuffer implements EditBuffer {
    private String text;
    ...
    public SimpleBuffer() {
        synchronized (this) {
            text = "";
            checkRep();
        }
    }
    public void insert(int pos, String ins) {
        synchronized (this) {
            text = text.substring(0, pos) + ins + text.substring(pos);
            checkRep();
        }
    }
    public void delete(int pos, int len) {
        synchronized (this) {
            text = text.substring(0, pos) + text.substring(pos+len);
            checkRep();
        }
    }
    public int length() {
        synchronized (this) {
            return text.length();
        }
    }
    public String toString() {
        synchronized (this) {
            return text;
        }
    }
}

Note the very careful discipline here. Every method that touches the rep must be guarded with the lock — even apparently small and trivial ones like length() and toString(). This is because reads must be guarded as well as writes — if reads are left unguarded, then they may be able to see the rep in a partially-modified state.

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.

Java provides some syntactic sugar that helps with the monitor pattern. If you add the keyword synchronized to a method signature, then Java will act as if you wrote synchronized (this) around the method body. So the code below is an equivalent way to implement the synchronized SimpleBuffer:

/** SimpleBuffer is a threadsafe EditBuffer with a simple rep. */
public class SimpleBuffer implements EditBuffer {
    private String text;
    ...
    public SimpleBuffer() {
        text = "";
        checkRep();
    }
    public synchronized void insert(int pos, String ins) {
        text = text.substring(0, pos) + ins + text.substring(pos);
        checkRep();
    }
    public synchronized void delete(int pos, int len) {
        text = text.substring(0, pos) + text.substring(pos+len);
        checkRep();
    }
    public synchronized int length() {
        return text.length();
    }
    public synchronized String toString() {
        return text;
    }
}

Notice that the SimpleBuffer constructor doesn’t have a synchronized keyword. Java actually forbids it, syntactically, because an object under construction is expected to be confined to a single thread until it has returned from its constructor. So synchronizing constructors should be unnecessary. You can still synchronize a constructor by wrapping its body in a synchronized(this) block, if you wish.

synchronized (obj) {
    // ...
    synchronized (obj) { // <-- uh oh, deadlock?
        // ...
    }
    // <-- do we own the lock on obj?
}

Now that we’re protecting SimpleBuffer’s rep with a lock, we can write a thread safety argument:

/** SimpleBuffer is a threadsafe EditBuffer with a simple rep. */
public class SimpleBuffer implements EditBuffer {
    private String text;
    // Rep invariant: 
    //   true
    // Abstraction function: 
    //   represents the sequence text[0],...,text[text.length()-1]
    // Safety from rep exposure:
    //   text is private and immutable
    // Thread safety argument:
    //   all accesses to text happen within SimpleBuffer methods,
    //   which are all guarded by SimpleBuffer's lock

The same argument works for GapBuffer, if we use the monitor pattern to synchronize all its methods.

Note that the encapsulation of the class, the absence of rep exposure, is very important for making this argument. If text were public:

    public String text;

then clients outside SimpleBuffer would be able to read and write it without knowing that they should first acquire the lock, and SimpleBuffer would no longer be threadsafe.

Locking discipline

Consider a find-and-replace operation on the EditBuffer datatype:

/** Modifies buf by replacing the first occurrence of s with t.
 *  If s not found in buf, then has no effect.
 *  @return true if and only if a replacement was made
 */
public static boolean findReplace(EditBuffer buf, String s, String t) {
    int i = buf.toString().indexOf(s);
    if (i == -1) {
        return false;
    }
    buf.delete(i, s.length());
    buf.insert(i, t);
    return true;
}

This method makes three different calls to buf — to convert it to a string in order to search for s, to delete the old text, and then to insert t in its place. Even though each of these calls individually is atomic, the findReplace method as a whole is not threadsafe, because other threads might mutate the buffer while findReplace is working, causing it to delete the wrong region or put the replacement back in the wrong place.

To prevent this, findReplace needs to synchronize with all other clients of buf.

Giving clients access to a lock

/** An EditBuffer represents a threadsafe mutable string of characters
 *  in a text editor. Clients may synchronize with each other using the
 *  EditBuffer object itself. */
public interface EditBuffer {
   ...
}

public static boolean findReplace(EditBuffer buf, String s, String t) {
    synchronized (buf) {
        int i = buf.toString().indexOf(s);
        if (i == -1) {
            return false;
        }
        buf.delete(i, s.length());
        buf.insert(i, t);
        return true;
    }
}

Sprinkling synchronized everywhere?

public static synchronized boolean findReplace(EditBuffer buf, ...) {

findReplace’s problem can be interpreted another way: that the EditBuffer interface really isn’t that friendly to multiple simultaneous clients. It relies on integer indexes to specify insert and delete locations, which are extremely brittle to other mutations. If somebody else inserts or deletes before the index position, then the index becomes invalid.

So if we’re designing a datatype specifically for use in a concurrent system, we need to think about providing operations that have better-defined semantics when they are interleaved. For example, it might be better to pair EditBuffer with a Position datatype representing a cursor position in the buffer, or even a Selection datatype representing a selected range. Once obtained, a Position could hold its location in the text against the wash of insertions and deletions around it, until the client was ready to use that Position. If some other thread deleted all the text around the Position, then the Position would be able to inform a subsequent client about what had happened (perhaps with an exception), and allow the client to decide what to do. These kinds of considerations come into play when designing a datatype for concurrency.

As another example, consider the ConcurrentMap interface in Java. This interface extends the existing Map interface, adding a few key methods that are commonly needed as atomic operations on a shared mutable map, e.g.:

• map.putIfAbsent(key,value) is an atomic version of
    if ( ! map.containsKey(key)) map.put(key, value);
• map.replace(key, value) is an atomic version of
    if (map.containsKey(key)) map.put(key, value);

The locking approach to thread safety is powerful, but (unlike confinement and immutability) it introduces blocking into the program. Threads must sometimes wait for other threads to get out of synchronized regions before they can proceed. And blocking raises the possibility of deadlock.

Deadlock happens when threads acquire multiple locks at the same time, and two threads end up blocked while holding locks that they are each waiting for the other to release. The monitor pattern unfortunately makes this fairly easy to do. Here’s an example.

Suppose we’re modeling the social network of a series of books:

public class Wizard {
    private final String name;
    private final Set<Wizard> friends;
    // Rep invariant:
    //    friend links are bidirectional: 
    //        for all f in friends, f.friends contains this
    // Concurrency argument:
    //    threadsafe by monitor pattern: all accesses to rep 
    //    are guarded by this object's lock

    public Wizard(String name) {
        this.name = name;
        this.friends = new HashSet<Wizard>();
    }

    public synchronized boolean isFriendsWith(Wizard that) {
        return this.friends.contains(that);
    }

    public synchronized void friend(Wizard that) {
        if (friends.add(that)) {
            that.friend(this);
        } 
    }

    public synchronized void defriend(Wizard that) {
        if (friends.remove(that)) {
            that.defriend(this);
        } 
    }
}

Like Facebook, this social network is bidirectional: if x is friends with y, then y is friends with x. The friend() and defriend() methods enforce that invariant by modifying the reps of both objects, which because they use the monitor pattern means acquiring the locks to both objects as well.

Let’s create a couple of wizards:

    Wizard harry = new Wizard("Harry Potter");
    Wizard snape = new Wizard("Severus Snape");

And then think about what happens when two independent threads are repeatedly running:

    // thread A                   // thread B
    harry.friend(snape);          snape.friend(harry);
    harry.defriend(snape);        snape.defriend(harry);

We will deadlock very rapidly. Here’s why. Suppose thread A is about to execute harry.friend(snape), and thread B is about to execute snape.friend(harry).

• Thread A acquires the lock on harry (because the friend method is synchronized).
• Then thread B acquires the lock on snape (for the same reason).
• They both update their individual reps independently, and then try to call friend() on the other object — which requires them to acquire the lock on the other object.

So A is holding Harry and waiting for Snape, and B is holding Snape and waiting for Harry. Both threads are stuck in friend(), so neither one will ever manage to exit the synchronized region and release the lock to the other. This is a classic deadly embrace. The program simply stops.

The essence of the problem is acquiring multiple locks, and holding some of the locks while waiting for another lock to become free.

Notice that it is possible for thread A and thread B to interleave such that deadlock does not occur: perhaps thread A acquires and releases both locks before thread B has enough time to acquire the first one. If the locks involved in a deadlock are also involved in a race condition — and very often they are — then the deadlock will be just as difficult to reproduce or debug.

Deadlock solution 1: lock ordering

    public void friend(Wizard that) {
        Wizard first, second;
        if (this.name.compareTo(that.name) < 0) {
            first = this; second = that;
        } else {
            first = that; second = this;
        }
        synchronized (first) {
            synchronized (second) {
                if (friends.add(that)) {
                    that.friend(this);
                } 
            }
        }
    }

Deadlock solution 2: coarse-grained locking

public class Wizard {
    private final Castle castle;
    private final String name;
    private final Set<Wizard> friends;
    ...
    public void friend(Wizard that) {
        synchronized (castle) {
            if (this.friends.add(that)) {
                that.friend(this);
            }
        }
    }
}

synchronized (alpha) {
    // using alpha
    // ...
}

synchronized (gamma) {
    synchronized (beta) {
        // using beta & gamma
        // ...
    }
}
// finished

synchronized (gamma) {
    synchronized (alpha) {
        synchronized (beta) {
            // using alpha, beta, & gamma
            // ...
        }
    }
}
// finished

synchronized (gamma) {
    synchronized (alpha) {
        // using alpha & gamma
        // ...
    }
}

synchronized (beta) {
    synchronized (gamma) {
        // using beta & gamma
        // ...
    }
}
// finished

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 threadsafe 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.

• Make thread safety arguments about your datatypes, and document them in the code.

• 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.