Can a good type system prevent concurrency errors? Or is this a quest for the Holy Grail?

There are two parts to this question, corresponding to two major types of concurrency errors:

  1. Preventing data races
  2. Preventing deadlocks

I’ll start with the first one.

Data races occur only when memory is shared between threads. Disallow sharing and data races are gone! In fact there is a name for threads that don’t share memory: processes. It’s perfectly feasible to have a concurrent language that disallows sharing–Erlang is one (see my post about Erlang). The trick is to always pass data between threads by value. This is especially easy in functional languages.

Non-functional languages like C++, Java, or D (I was told by Walter Bright, the creator of D, to always use the full name, “the D programming language,” so that search engines can index it properly) tend to share data almost by default (see, however, this post).

In Java, all non-primitive types have reference semantics. When you pass a Java object between threads, you’re only passing a handle; the object behind it becomes accessible from both threads.

C++ at least has means to express passing by value and move semantics for user-defined types. Still, it’s up to the programmer to use them.

Who ordered Guava?

For every type-system idea there is one or more dialects of Java that implement it. I’ll start with an older attempt at data-race free Java called Guava, as it illustrates some of the basic premises.

-Explicit sharing

The most important step–if we don’t want to completely ban the sharing of data–is to regulate it. Let the programmer explicitly mark the data destined for sharing as such. The corollary is that the data that is not marked for sharing cannot be shared. This can be accomplished, for instance, by making all objects thread-local by default, or by using type modifiers that prevent references to such objects from escaping.

In Guava, the data type designed for sharing is called a Monitor. As the name suggests, all access to a Monitor is automatically synchronized by the Monitor’s lock. This, incidentally, eliminates the need for the synchronized keyword, which is absent from Guava.

The non-shared data types are either Objects or Values.

Objects are just like regular Java Objects, except that they don’t have a built-in lock, since they can never be shared between threads.

Values are either primitive values, like integers, or user-defined types with value semantics.

Monitors, Objects, and Values are collectively called instances.

-Value semantics

When you pass a Value, the compiler will make a deep copy of it (well, sort of–the monitors embedded in a Value are not deep copied). Since deep copying might be expensive, Guava defines operator “move”, which nulls the source. The syntax is:

  v2 = v1--;

The value v1 becomes null after the move to v2. This is similar to C++ unique_ptr and std::move.


The biggest problem in lock based concurrency is to make sure that the correct lock(s) are taken when accessing shared data. In Guava, all Monitor’s data are protected by that Monitor’s lock. As long as they stay inside that Monitor, nothing bad can happen to them from the point of concurrency.

Values stored inside a Monitor are never accessible outside of the Monitor–only their copies may be passed out.

The same is not true about Objects. Since Objects have reference semantics, there is a real danger that Objects’ references might escape the Monitor that protects them. Imagine a situation where two Monitors have references to the same Object. It is possible then that two threads may operate on that Object at the same time–one entering through one Monitor, the other through the other Monitor. We have a data race!

Therefore it is important that every Object have an owner at all times. The Object’s owner is either a Value or a Monitor. (The exception is a fresh Object that’s been just allocated–it has no owner until it is assigned to a variable.) Since an Object may only be owned by at most one Monitor, it is that Monitor that protects it from simultaneous (racy) access.


All Objects that are owned by a particular Monitor or Value form a region. Equivalently, assigning a monitored region to an object specifies what lock must be held when accessing it.

All instances may contain (references to) monitors, but monitors are not “owned” by anybody. References to the same monitor may appear in multiple regions and may be freely passed around. It is thus up to programmers to define an ordering scheme for their monitors in order to avoid deadlocks.

How can we protect Objects from moving between regions and acquiring multiple owners? We need a way to control aliasing.

Here are some Guava rules for passing Objects. A method may declare its Object parameter as either kept or lent. (By default, parameters to Object methods are kept and to Monitor methods are lent.) If the parameter is kept it must belong to the same region as this, and there are no limitations on its use. If, however, the parameter is lent, the method may not store a reference to it in this, nor may it store this inside a lent Object. No cross-aliasing is allowed.

A method may also be marked new if it returns a freshly created object, which has no owner yet. Constructors are considered new unless they accept kept parameters.

Notice that you may have multiple references to the same Object, but they will all be within the same region. The only instances that may be passed between threads are Monitors and Values.


Guava final fields may either be initialized inside a constructor or in a private method that is only called from inside a constructor. (BTW, in D a private method is not private within a module, so the compiler would have to analyze the whole module to establish the latter condition.) [Note to self: The same scheme might be useful in the construction of immutable objects in D.]

Partially constructed Objects must not be visible outside constructors. The compiler must verify that constructors don’t pass this to other methods, and don’t store this inside other instances (no alias cross-contamination).


Copying Values around may be expensive. I already mentioned one optimization, the use of the “move” operator. The other optimization is related to immutability. If a Value is immutable, it may be safely passed by reference. Guava defines immutable classes as ones that have no update methods. Any method that may modify this must be marked update. The update notation is also extended to method parameters–by default parameters are immutable.

There is a bonus advantage to separating update methods from the rest. In a Monitor, a non-update method may safely use a readers’ lock, which allows multiple readers to access data simultaneously, to increase concurrency.

-Global and local methods

A method is considered local if its effects cannot be observed or influenced by other threads. All Object and Value methods are by default local. A local method is immune to races thus allowing single-thread optimizations.

Conversely, all Monitor methods are considered global, since operations on Monitors may be visible or influenced by other threads.

These defaults may be overridden. For instance, an Object may contain a reference to a Monitor. The methods that call this Monitor’s methods must be marked global. Moreover, Object or Value methods that access Monitors that are passed to them as arguments must also be marked global. So touching a Monitor (which is equivalent to using its lock) taints the whole callers’ tree with the global annotation.

This is similar to the way update taints the callers tree, except the update annotation of a method only pertains to this, not to its arguments. However, when global and update are combined, they have the same tainting power as global. In particular, a method that invokes a global update method on its argument becomes tainted with global update.

Methods that are global update cannot be invoked from non-update methods, even if they only global-update their arguments.

Note: This part of the paper is not very clear to me. The authors never explain the importance of global update methods (other than optimization opportunities).


Guava implements a relatively simple and somewhat limited system to eliminate races. It punts the problem of ownership-passing and lock-free programming. Even with those limitations the Guava type system is not simple.

The idea that safe multithreading may be achieved with a few simple modification to the type system seems to be untenable. However, as long as special type annotations are limited to the code that does actual multithreading, I think they’re worth the added complexity.