Mutex (abbreviated Mutually Exclusive access) is a synchronization object, a variant of semaphore with k=1. Mutex is said to be seized by a task decreasing k. It is released when the task restores k. Mutexes are typically used to protect a shared resource from concurrent access. A task seizes the mutex, then accesses the resource, and after that releases the mutex.

Mutex is a low-level synchronization primitive exposed to deadlocking. A deadlock can occur with just two tasks and two mutexes (if each task attempts to acquire both mutexes, but in the opposite order). Entering the deadlock is usually aggravated by a race condition state, which leads to sporadic hangups, very difficult to track down.

Variants of mutexes

Global and local mutexes

Usually OS provides various implementations of mutexes corresponding to the variants of tasks available in the OS. For example, system-wide mutexes can be used by processes. Local mutexes can be used only by threads etc. This distinction is maintained because, depending on the hardware, seizing a global mutex might be thousand times slower than seizing a local one.

Reentrant mutex

A reentrant mutex can be seized by the same task multiple times. Each seizing of the mutex is matched by releasing it, in order to allow other task to seize it.

Read write mutex

A read write mutex can be seized at two levels for read and for write. The mutex can be seized for read by any number of tasks. Only one task may seize it 'write. Read write mutex are usually used to protect resources which can be accessed in a mutable and immutable ways. Immutable (read) access is granted concurrently for many tasks because they do not change the resource state. Read write mutexes can be reentrant, global or local. Further, promotion operations may be provided, that's when a task that have seized the mutex for write releases it keeping seized for read. Note that the reverse operation is potentially deadlocking and requires some additional access policy control.

Deadlock prevention

There exists a simple technique of deadlock prevention when mutexes are seized in some fixed order. This is discussed in depth in the Dining philosophers problem.

Sample implementations / APIs

Ada

Ada provides higher-level concurrency primitives, which are complete in the sense that they also allow implementations of the lower-level ones, like mutex. Here is an implementation of a plain non-reentrant mutex based on protected objects.

The mutex interface: <lang ada> protected type Mutex is

  entry Seize;
  procedure Release;

private

  Owned : Boolean := False;

end Mutex; </lang> The implementation of: <lang ada> protected body Mutex is

  entry Seize when not Owned is
  begin
     Owned := True;
  end Seize;
  procedure Release is
  begin
     Owned := False;
  end Release;

end Mutex; </lang> Here the entry Seize has a queue of the tasks waiting for the mutex. The entry's barrier is closed when Owned is true. So any task calling to the entry will be queued. When the barrier is open the first task from the queue executes the entry and Owned becomes false closing the barrier again. The procedure Release simply sets Owned to false. Both Seize and Release are protected actions which execution causes reevaluation of all barriers, in this case one of Seize.

Use: <lang ada> declare

  M : Mutex;

begin

  M.Seize;    -- Wait infinitely for the mutex to be free
  ...         -- Critical code
  M.Release;  -- Release the mutex
  ...
  select
     M.Seize; -- Wait no longer than 0.5s
  or delay 0.5;
     raise Timed_Out;
  end select;
  ...         -- Critical code
  M.Release;  -- Release the mutex

end; </lang> It is also possible to implement mutex as a monitor task.

C

Win32

Works with: Win32

To create a mutex operating system "object": <lang c>

 HANDLE hMutex = CreateMutex(NULL, FALSE, NULL);

</lang> To lock the mutex: <lang c>

 WaitForSingleObject(hMutex, INFINITE);

</lang> To unlock the mutex <lang c>

 ReleaseMutex(hMutex);

</lang> When the program is finished with the mutex: <lang c>

 CloseHandle(hMutex);

</lang>

POSIX

Works with: POSIX

Creating a mutex:

<lang c>#include <pthread.h>

pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;</lang>

Or:

<lang c>pthread_mutex_t mutex; pthread_mutex_init(&mutex, NULL);</lang>

Locking:

<lang c>int success = pthread_mutex_lock(&mutex);</lang>

Unlocking:

<lang c>int success = pthread_mutex_unlock(&mutex);</lang>

Trying to lock (but do not wait if it can't)

<lang c>int success = pthread_mutex_trylock(&mutex);</lang>

C++

Works with: Win32

See C example

E

E's approach to concurrency is to never block, in favor of message passing/event queues/callbacks. Therefore, it is unidiomatic to use a mutex at all, and incorrect, or rather unsafe, to use a mutex which blocks the calling thread. That said, here is a mutex written in E.

<lang e>def makeMutex() {

   # The mutex is available (released) if available is resolved, otherwise it
   # has been seized/locked. The specific value of available is irrelevant.
   var available := null
   
   # The interface to the mutex is a function, taking a function (action)
   # to be executed. 
   def mutex(action) {
       # By assigning available to our promise here, the mutex remains 
       # unavailable to the /next/ caller until /this/ action has gotten
       # its turn /and/ resolved its returned value.
       available := Ref.whenResolved(available, fn _ { action <- () })
   }
   return mutex

}</lang>

This implementation of a mutex is designed to have a very short implementation as well as usage in E. The mutex object is a function which takes a function action to be executed once the mutex is available. The mutex is unavailable until the return value of action resolves. This interface has been chosen over lock and unlock operations to reduce the hazard of unbalanced lock/unlock pairs, and because it naturally fits into E code.

Usage example:

<lang e>Creating the mutex:

? def mutex := makeMutex()

value: <mutex>

Creating the shared resource:

? var value := 0

value: 0

Manipulating the shared resource non-atomically so as to show a problem:

? for _ in 0..1 { > when (def v := (&value) <- get()) -> { > (&value) <- put(v + 1) > } > }

? value

value: 1

The value has been incremented twice, but non-atomically, and so is 1 rather than the intended 2.

? value := 0

value: 0

This time, we use the mutex to protect the action.

? for _ in 0..1 { > mutex(fn { > when (def v := (&value) <- get()) -> { > (&value) <- put(v + 1) > } > }) > }

? value

value: 2</lang>

when blocks and Ref.whenResolved return a promise for the result of the deferred action, so the mutex here waits for the gratuitously complicated increment to complete before becoming available for the next action.

Objective-C

<lang objc>NSLock *m = [[NSLock alloc] init];

[m lock]; // locks in blocking mode

if ([m tryLock]) { // acquire a lock -- does not block if not acquired

 // lock acquired

} else {

 // already locked, does not block

}

[m unlock];</lang>

Reentrant mutex is provided by the NSRecursiveLock class.

OCaml

OCaml provides a built-in Mutex module.
It is very simple, there are four functions:

<lang ocaml>let m = Mutex.create() in Mutex.lock m; (* locks in blocking mode *)

if (Mutex.try_lock m) then ... (* did the lock *) else ... (* already locked, do not block *)

Mutex.unlock m;</lang>

Ruby

Ruby's standard library includes a mutex_m module that can be mixed-in to a class. <lang ruby>require 'mutex_m'

class SomethingWithMutex

 include Mutex_m
 ...

end</lang>

Individual objects can be extended with the module too <lang ruby>an_object = Object.new an_object.extend(Mutex_m)</lang>

An object with mutex powers can then: <lang ruby># acquire a lock -- block execution until it becomes free an_object.mu_lock

acquire a lock -- return immediately even if not acquired

got_lock = an_object.mu_try_lock

have a lock?

if an_object.mu_locked? then ...

release the lock

an_object.mu_unlock

wrap a lock around a block of code -- block execution until it becomes free

an_object.my_synchronize do

 do critical stuff

end</lang>

Tcl

Tcl's mutexes have four functions. <lang tcl>package require Thread

How to create a mutex

set m [thread::mutex create]

This will block if the lock is already held unless the mutex is made recursive

thread::mutex lock $m

Now locked...

thread::mutex unlock $m

Unlocked again

Dispose of the mutex

thread::mutex destroy $m</lang> There are also read-write mutexes available. <lang tcl>set rw [thread::rwmutex create]

Get and drop a reader lock

thread::rwmutex rlock $rw thread::rwmutex unlock $rw

Get and drop a writer lock

thread::rwmutex wlock $rw thread::rwmutex unlock $rw

thread::rwmutex destroy $rw</lang>