The concurrency extension for Haskell is described in the paper Concurrent Haskell http://www.haskell.org/ghc/docs/papers/concurrent-haskell.ps.gz.

Concurrency is "lightweight", which means that both thread creation and context switching overheads are extremely low. Scheduling of Haskell threads is done internally in the Haskell runtime system, and doesn't make use of any operating system-supplied thread packages.

However, if you want to interact with a foreign library that expects your program to use the operating system-supplied thread package, you can do so by using forkOS instead of forkIO.

Haskell threads can communicate via MVars, a kind of synchronised mutable variable (see Control.Concurrent.MVar). Several common concurrency abstractions can be built from MVars, and these are provided by the Control.Concurrent library. In GHC, threads may also communicate via exceptions.

This sparks off a new thread to run the IO computation passed as the first argument, and returns the ThreadId of the newly created thread.

The new thread will be a lightweight thread; if you want to use a foreign library that uses thread-local storage, use forkOS instead.

killThread terminates the given thread (GHC only). Any work already done by the thread isn't lost: the computation is suspended until required by another thread. The memory used by the thread will be garbage collected if it isn't referenced from anywhere. The killThread function is defined in terms of throwTo:

 killThread tid = throwTo tid (AsyncException ThreadKilled)

throwTo raises an arbitrary exception in the target thread (GHC only).

throwTo does not return until the exception has been raised in the target thread. The calling thread can thus be certain that the target thread has received the exception. This is a useful property to know when dealing with race conditions: eg. if there are two threads that can kill each other, it is guaranteed that only one of the threads will get to kill the other.

If the target thread is currently making a foreign call, then the exception will not be raised (and hence throwTo will not return) until the call has completed. This is the case regardless of whether the call is inside a block or not.

Important note: the behaviour of throwTo differs from that described in the paper Asynchronous exceptions in Haskell (http://research.microsoft.com/~simonpj/Papers/asynch-exns.htm). In the paper, throwTo is non-blocking; but the library implementation adopts a more synchronous design in which throwTo does not return until the exception is received by the target thread. The trade-off is discussed in Section 8 of the paper. Like any blocking operation, throwTo is therefore interruptible (see Section 4.3 of the paper).

There is currently no guarantee that the exception delivered by throwTo will be delivered at the first possible opportunity. In particular, if a thread may unblock and then re-block exceptions (using unblock and block) without receiving a pending throwTo. This is arguably undesirable behaviour.

Scheduling may be either pre-emptive or co-operative, depending on the implementation of Concurrent Haskell (see below for information related to specific compilers). In a co-operative system, context switches only occur when you use one of the primitives defined in this module. This means that programs such as:

   main = forkIO (write 'a') >> write 'b'
     where write c = putChar c >> write c

will print either aaaaaaaaaaaaaa... or bbbbbbbbbbbb..., instead of some random interleaving of as and bs. In practice, cooperative multitasking is sufficient for writing simple graphical user interfaces.

Different Haskell implementations have different characteristics with regard to which operations block all threads.

Using GHC without the -threaded option, all foreign calls will block all other Haskell threads in the system, although I/O operations will not. With the -threaded option, only foreign calls with the unsafe attribute will block all other threads.

Using Hugs, all I/O operations and foreign calls will block all other Haskell threads.

Suspends the current thread for a given number of microseconds (GHC only).

There is no guarantee that the thread will be rescheduled promptly when the delay has expired, but the thread will never continue to run earlier than specified.

The mergeIO and nmergeIO functions fork one thread for each input list that concurrently evaluates that list; the results are merged into a single output list.

Note: Hugs does not provide these functions, since they require preemptive multitasking.

Support for multiple operating system threads and bound threads as described below is currently only available in the GHC runtime system if you use the -threaded option when linking.

Other Haskell systems do not currently support multiple operating system threads.

A bound thread is a haskell thread that is bound to an operating system thread. While the bound thread is still scheduled by the Haskell run-time system, the operating system thread takes care of all the foreign calls made by the bound thread.

To a foreign library, the bound thread will look exactly like an ordinary operating system thread created using OS functions like pthread_create or CreateThread.

Bound threads can be created using the forkOS function below. All foreign exported functions are run in a bound thread (bound to the OS thread that called the function). Also, the main action of every Haskell program is run in a bound thread.

Why do we need this? Because if a foreign library is called from a thread created using forkIO, it won't have access to any thread-local state - state variables that have specific values for each OS thread (see POSIX's pthread_key_create or Win32's TlsAlloc). Therefore, some libraries (OpenGL, for example) will not work from a thread created using forkIO. They work fine in threads created using forkOS or when called from main or from a foreign export.

Like forkIO, this sparks off a new thread to run the IO computation passed as the first argument, and returns the ThreadId of the newly created thread.

However, forkOS uses operating system-supplied multithreading support to create a new operating system thread. The new thread is bound, which means that all foreign calls made by the IO computation are guaranteed to be executed in this new operating system thread; also, the operating system thread is not used for any other foreign calls.

This means that you can use all kinds of foreign libraries from this thread (even those that rely on thread-local state), without the limitations of forkIO.

Just to clarify, forkOS is only necessary if you need to associate a Haskell thread with a particular OS thread. It is not necessary if you only need to make non-blocking foreign calls (see Control.Concurrent#osthreads). Neither is it necessary if you want to run threads in parallel on a multiprocessor: threads created with forkIO will be shared out amongst the running CPUs (using GHC, -threaded, and the +RTS -N runtime option).

Run the IO computation passed as the first argument. If the calling thread is not bound, a bound thread is created temporarily. runInBoundThread doesn't finish until the IO computation finishes.

You can wrap a series of foreign function calls that rely on thread-local state with runInBoundThread so that you can use them without knowing whether the current thread is bound.

Run the IO computation passed as the first argument. If the calling thread is bound, an unbound thread is created temporarily using forkIO. runInBoundThread doesn't finish until the IO computation finishes.

Use this function only in the rare case that you have actually observed a performance loss due to the use of bound threads. A program that doesn't need it's main thread to be bound and makes heavy use of concurrency (e.g. a web server), might want to wrap it's main action in runInUnboundThread.

In GHC, threads created by forkIO are lightweight threads, and are managed entirely by the GHC runtime. Typically Haskell threads are an order of magnitude or two more efficient (in terms of both time and space) than operating system threads.

The downside of having lightweight threads is that only one can run at a time, so if one thread blocks in a foreign call, for example, the other threads cannot continue. The GHC runtime works around this by making use of full OS threads where necessary. When the program is built with the -threaded option (to link against the multithreaded version of the runtime), a thread making a safe foreign call will not block the other threads in the system; another OS thread will take over running Haskell threads until the original call returns. The runtime maintains a pool of these worker threads so that multiple Haskell threads can be involved in external calls simultaneously.

The System.IO library manages multiplexing in its own way. On Windows systems it uses safe foreign calls to ensure that threads doing I/O operations don't block the whole runtime, whereas on Unix systems all the currently blocked I/O reqwests are managed by a single thread (the IO manager thread) using select.

The runtime will run a Haskell thread using any of the available worker OS threads. If you need control over which particular OS thread is used to run a given Haskell thread, perhaps because you need to call a foreign library that uses OS-thread-local state, then you need bound threads (see Control.Concurrent#boundthreads).

If you don't use the -threaded option, then the runtime does not make use of multiple OS threads. Foreign calls will block all other running Haskell threads until the call returns. The System.IO library still does multiplexing, so there can be multiple threads doing I/O, and this is handled internally by the runtime using select.

In a standalone GHC program, only the main thread is required to terminate in order for the process to terminate. Thus all other forked threads will simply terminate at the same time as the main thread (the terminology for this kind of behaviour is "daemonic threads").

If you want the program to wait for child threads to finish before exiting, you need to program this yourself. A simple mechanism is to have each child thread write to an MVar when it completes, and have the main thread wait on all the MVars before exiting:

   myForkIO :: IO () -> IO (MVar ())
   myForkIO io = do
     mvar <- newEmptyMVar
     forkIO (io `finally` putMVar mvar ())
     return mvar

Note that we use finally from the Control.Exception module to make sure that the MVar is written to even if the thread dies or is killed for some reason.

A better method is to keep a global list of all child threads which we should wait for at the end of the program:

    children :: MVar [MVar ()]
    children = unsafePerformIO (newMVar [])
    
    waitForChildren :: IO ()
    waitForChildren = do
      cs <- takeMVar children
      case cs of
        []   -> return ()
        m:ms -> do
    	    putMVar children ms
    	    takeMVar m
    	    waitForChildren
    
    forkChild :: IO () -> IO ThreadId
    forkChild io = do
    	 mvar <- newEmptyMVar
    	 childs <- takeMVar children
    	 putMVar children (mvar:childs)
    	 forkIO (io `finally` putMVar mvar ())

     main =
     	later waitForChildren $
     	...

The main thread principle also applies to calls to Haskell from outside, using foreign export. When the foreign exported function is invoked, it starts a new main thread, and it returns when this main thread terminates. If the call causes new threads to be forked, they may remain in the system after the foreign exported function has returned.

GHC implements pre-emptive multitasking: the execution of threads are interleaved in a random fashion. More specifically, a thread may be pre-empted whenever it allocates some memory, which unfortunately means that tight loops which do no allocation tend to lock out other threads (this only seems to happen with pathological benchmark-style code, however).

The rescheduling timer runs on a 20ms granularity by default, but this may be altered using the -i<n> RTS option. After a rescheduling "tick" the running thread is pre-empted as soon as possible.

One final note: the aaaa bbbb example may not work too well on GHC (see Scheduling, above), due to the locking on a Handle. Only one thread may hold the lock on a Handle at any one time, so if a reschedule happens while a thread is holding the lock, the other thread won't be able to run. The upshot is that the switch from aaaa to bbbbb happens infrequently. It can be improved by lowering the reschedule tick period. We also have a patch that causes a reschedule whenever a thread waiting on a lock is woken up, but haven't found it to be useful for anything other than this example :-)