Initial revision

[platform/upstream/glib.git] / docs / reference / glib / tmpl / threads.sgml

<!-- ##### SECTION Title ##### -->

Threads

<!-- ##### SECTION Short_Description ##### -->

thread abstraction; including mutexes, conditions and thread private data.

<!-- ##### SECTION Long_Description ##### -->

<para>
Threads act almost like processes, but unlike processes all threads of
one process share the same memory. This is good, as it provides easy
communication between the involved threads via this shared memory, and
it is bad, because strange things (so called Heisenbugs) might happen,
when the program is not carefully designed. Especially bad is, that due
to the concurrent nature of threads no assumptions on the order of
execution of different threads can be done unless explictly forced by
the programmer through synchronization primitives.
</para>

<para>
The aim of the thread related functions in GLib is to provide a
portable means for writing multithread safe software. There are
primitives for mutexes to protect the access to portions of memory
(#GMutex, #GStaticMutex, #G_LOCK_DEFINE and friends), there are
primitives for condition variables to allow synchronization of threads
(#GCond) and finally there are primitives for thread-private data,
that every thread has a private instance of (#GPrivate,
#GStaticPrivate).
</para>

<para>
Currently there is only as much thread support included in GLib as is
necessary to make GLib itself multithread safe. Future versions of
GLib might contain functions to actually create threads and the
like. For now the most portable way to create threads is to require
the macro #G_THREADS_IMPL_POSIX to be defined and use POSIX threads
then. This will work on almost all platforms (except most notably
Solaris).
</para>

<!-- ##### SECTION See_Also ##### -->
<para>

</para>

<!-- ##### MACRO G_THREADS_ENABLED ##### -->

<para>
This macro is defined, if GLib was compiled with thread support. This
does not necessarily mean, that there is a thread implementation
available, but the infrastructure is in place and once you provide a
thread implementation to g_thread_init(), GLib will be multithread
safe. It isn't and can't be, if #G_THREADS_ENABLED is not defined.
</para>



<!-- ##### MACRO G_THREADS_IMPL_POSIX ##### -->

<para>
This macro is defined, if POSIX style threads are used.
</para>



<!-- ##### MACRO G_THREADS_IMPL_SOLARIS ##### -->

<para>
This macro is defined, if the SOLARIS thread system is used.
</para>



<!-- ##### MACRO G_THREADS_IMPL_NSPR ##### -->

<para>
This macro is defined, if the NSPR thread implementation is used.
NSPR is the cross platform library of mozilla.
</para>



<!-- ##### MACRO G_THREADS_IMPL_NONE ##### -->

<para>
This macro is defined, if no thread implementation is used. You can
however provide one to g_thread_init() to make GLib multithread safe.
</para>



<!-- ##### STRUCT GThreadFunctions ##### -->

<para>
This function table is used by g_thread_init() to initialize the
thread system. The functions in that table are directly used by their
g_* prepended counterparts, that are described here, e.g. if you call
g_mutex_new() then mutex_new() from the table provided to
g_thread_init() will be called.
</para>

<note>
<para>
This struct should only be used, if you know, what you are doing.
</para>
</note>

@mutex_new: 
@mutex_lock: 
@mutex_trylock: 
@mutex_unlock: 
@mutex_free: 
@cond_new: 
@cond_signal: 
@cond_broadcast: 
@cond_wait: 
@cond_timed_wait: 
@cond_free: 
@private_new: 
@private_get: 
@private_set: 

<!-- ##### FUNCTION g_thread_init ##### -->

<para>
Before you use a thread related function in GLib, you should
initialize the thread system. This is done by calling
g_thread_init(). Most of the time you will only have to call
g_thread_init(NULL). 
</para>

<note>
<para>
You should only call g_thread_init() with a non-NULL parameter, if you
really know, what you are doing.
</para>
</note>

<note>
<para>
g_thread_init() must not be called directly or indirectly as a
callback from GLib.
</para>
</note>

<para>
g_thread_init() might only be called once. On the second call
it will abort with an error. If you want to make sure, that the thread
system is initialized, you can do that too:
</para>

<para>
<informalexample>
<programlisting>
if (!g_thread_supported ()) g_thread_init (NULL);
</programlisting>
</informalexample>
</para>

<para>
After that line either the thread system is initialized or the program
will abort, if no thread system is available in GLib, i.e. either
#G_THREADS_ENABLED is not defined or #G_THREADS_IMPL_NONE is defined.
</para>

<para>
If no thread system is available and @vtable is NULL or if not all
elements of @vtable are non-NULL, then g_thread_init() will abort.
</para>

<note>
<para>
To use g_thread_init() in your program, you have to link with the
libraries, that the command "glib-config --libs gthread" outputs. This
is not the case for all the other thread related functions of
GLib. Those can be used without having to link with the thread
libraries.
</para>
</note>

@vtable: a function table of type #GThreadFunctions, that provides the
entry points to the thread system to be used.


<!-- ##### MACRO g_thread_supported ##### -->
<para>
This function returns, whether the thread system is initialized or
not.
</para>

<note>
<para>
This function is actually a macro. Apart from taking the address of it
you can however use it as if it was a function.
</para>
</note>

@Returns: TRUE, if the thread system is initialized.


<!-- ##### STRUCT GMutex ##### -->

<para>
The #GMutex struct is an opaque data structure to represent a mutex
(mutual exclusion). It can be used to protect data against shared
access. Take for example the following function:

<para>
<example>
<title>A function which will not work in a threaded environment</title>
<programlisting>
  int give_me_next_number ()
  {
    static int current_number = 0;

    /* now do a very complicated calculation to calculate the new number,
       this might for example be a random number generator */
    current_number = calc_next_number (current_number); 
    return current_number;
  }
</programlisting>
</example>
</para>

<para>
It is easy to see, that this won't work in a multithreaded
application. There current_number must be protected against shared
access. A first naive implementation would be:
</para>

<para>
<example>
<title>The wrong way to write a thread-safe function</title>
<programlisting>
  int give_me_next_number ()
  {
    static int current_number = 0;
    int ret_val;
    static GMutex * mutex = NULL;

    if (!mutex)
      mutex = g_mutex_new ();
    g_mutex_lock (mutex);
    ret_val = current_number = calc_next_number (current_number); 
    g_mutex_unlock (mutex);
    return ret_val;
  }
</programlisting>
</example>
</para>

<para>
This looks like it would work, but there is a race condition while
constructing the mutex and this code can't work reliable. So please do
not use such constructs in your own programs. One working solution is:
</para>

<para>
<example>
<title>A correct thread-safe function</title>
<programlisting>
  static GMutex *give_me_next_number_mutex = NULL;

  /* this function must be called before any call to give_me_next_number ()
     it must be called exactly once. */
  void init_give_me_next_number () 
  {
    g_assert (give_me_next_number_mutex == NULL);
    give_me_next_number_mutex = g_mutex_new ();
  }

  int give_me_next_number ()
  {
    static int current_number = 0;
    int ret_val;

    g_mutex_lock (give_me_next_number_mutex);
    ret_val = current_number = calc_next_number (current_number); 
    g_mutex_unlock (give_me_next_number_mutex);
    return ret_val;
  }
</programlisting>
</example>
</para>

<para>
#GStaticMutex provides a simpler and safer way of doing this.
</para>

<para>
A #GMutex should only be accessed via the following functions.
</para>

<note>
<para>
All of the g_mutex_* functions are actually macros. Apart from taking
the addresses of them, you can however use them as if they were functions.
</para>
</note>


<!-- ##### MACRO g_mutex_new ##### -->

<para>
Creates a new #GMutex. 
</para>

<note>
<para>
This function will abort, if g_thread_init() has not been called yet.
</para>
</note>

@Returns: a new #GMutex.


<!-- ##### MACRO g_mutex_lock ##### -->

<para>
Locks the #GMutex. If the #GMutex is already locked by another thread,
the current thread will block until the #GMutex is unlocked by the
other thread.
</para>

<para>
This function can also be used, if g_thread_init() has not yet been
called and will do nothing then.
</para>

<note>
<para>
#GMutex is not guaranteed to be recursive, i.e. a thread might block,
if it already has locked the #GMutex. It will deadlock then, of
course.
</para>
</note>

@mutex: a #GMutex.


<!-- ##### MACRO g_mutex_trylock ##### -->

<para>
Tries to lock the #GMutex. If the #GMutex is already locked by another
thread, it immediately returns FALSE. Otherwise it locks the #GMutex
and returns TRUE.
</para>

<para>
This function can also be used, if g_thread_init() has not yet been
called and will immediately return TRUE then.
</para>

@mutex: a #GMutex.
@Returns: TRUE, if the #GMutex could be locked.


<!-- ##### MACRO g_mutex_unlock ##### -->

<para>
Unlocks the #GMutex. If another thread is blocked in a g_mutex_lock()
call, it will be woken and can lock the #GMutex itself. This function
can also be used, if g_thread_init() has not yet been called and will
do nothing then.
</para>

@mutex: a #GMutex.


<!-- ##### MACRO g_mutex_free ##### -->

<para>
Destroys the #GMutex.
</para>

@mutex: a #GMutex.


<!-- ##### STRUCT GStaticMutex ##### -->

<para>
A #GStaticMutex works like a #GMutex, but it has one significant
advantage. It doesn't need to be created at run-time like a #GMutex,
but can be defined at compile-time. Here is a shorter, easier and
safer version of our give_me_next_number() example:
</para>

<para>
<example>
<title>Using GStaticMutex to simplify thread-safe programming</title>
<programlisting>
  int give_me_next_number ()
  {
    static int current_number = 0;
    int ret_val;
    static GStaticMutex mutex = G_STATIC_MUTEX_INIT;

    g_static_mutex_lock (&amp;mutex);
    ret_val = current_number = calc_next_number (current_number); 
    g_static_mutex_unlock (&amp;mutex);
    return ret_val;
  }
</programlisting>
</example>
</para>

<para>
Even though #GStaticMutex is not opaque, it should only be used with
the following functions, as it is defined differently on different
platforms.
</para>

<para>All of the g_static_mutex_* functions can also be used, if
g_thread_init() has not yet.
</para>

<note>
<para>
All of the g_static_mutex_* functions are actually macros. Apart from
taking the addresses of them, you can however use them as if they were
functions.
</para>
</note>

@runtime_mutex: 

<!-- ##### MACRO G_STATIC_MUTEX_INIT ##### -->

<para>
Every #GStaticMutex must be initialized with this macro, before it can
be used.
</para>

<para>
<example>
<title>Initializing a GStaticMutext</title>
<programlisting>
GStaticMutex my_mutex = G_STATIC_MUTEX_INIT;
</programlisting>
</example>
</para>



<!-- ##### MACRO g_static_mutex_lock ##### -->
<para>
works like g_mutex_lock(), but for a #GStaticMutex.
</para>

@mutex: a #GStaticMutex.


<!-- ##### MACRO g_static_mutex_trylock ##### -->

<para>
works like g_mutex_trylock(), but for a #GStaticMutex.
</para>

@mutex: a #GStaticMutex.
@Returns: TRUE, if the #GStaticMutex could be locked.


<!-- ##### MACRO g_static_mutex_unlock ##### -->

<para>
works like g_mutex_unlock(), but for a #GStaticMutex.
</para>

@mutex: a #GStaticMutex.


<!-- ##### MACRO g_static_mutex_get_mutex ##### -->

<para>
For some operations (like g_cond_wait()) you must have a #GMutex
instead of a #GStaticMutex. This function will return the
corresponding #GMutex for every #GStaticMutex.
</para>

@mutex: a #GStaticMutex.
@Returns: the corresponding #GMutex.


<!-- ##### MACRO G_LOCK_DEFINE ##### -->

<para>
The G_LOCK_* macros provide a convenient interface to #GStaticMutex
with the advantage that they will expand to nothing in programs
compiled against a thread-disabled GLib, saving code and memory
there. #G_LOCK_DEFINE defines a lock. It can occur, where variable
definitions may occur in programs, i.e. in the first block of a
function or outside of functions. The @name parameter will be mangled
to get the name of the #GStaticMutex. This means, that you can use
names of existing variables as the parameter, e.g. the name of the
variable you intent to protect with the lock. Look at our
give_me_next_number() example using the G_LOCK_* macros:
</para>

<para>
<example>
<title>Using the G_LOCK_* convenience macros</title>
<programlisting>
int give_me_next_number ()
  {
    static int current_number = 0;
    int ret_val;
    G_LOCK_DEFINE_STATIC (current_number);

    G_LOCK (current_number);
    ret_val = current_number = calc_next_number (current_number); 
    G_UNLOCK (current_number);
    return ret_val;
  }
</programlisting>
</example>
</para>

@name: the name of the lock.


<!-- ##### MACRO G_LOCK_DEFINE_STATIC ##### -->

<para>
This works like #G_LOCK_DEFINE, but it creates a static object.
</para>

@name: the name of the lock.


<!-- ##### MACRO G_LOCK_EXTERN ##### -->

<para>
This declares a lock, that is defined with #G_LOCK_DEFINE in another module.
</para>

@name: the name of the lock.


<!-- ##### MACRO G_LOCK ##### -->

<para>
works like g_mutex_lock(), but for a lock defined with #G_LOCK_DEFINE.
</para>

@name: the name of the lock.


<!-- ##### MACRO G_TRYLOCK ##### -->

<para>
works like g_mutex_trylock(), but for a lock defined with #G_LOCK_DEFINE.
</para>

@name: the name of the lock.
@Returns: TRUE, if the lock could be locked.


<!-- ##### MACRO G_UNLOCK ##### -->

<para>
works like g_mutex_unlock(), but for a lock defined with #G_LOCK_DEFINE.
</para>

@name: the name of the lock.


<!-- ##### STRUCT GCond ##### -->

<para>
The #GCond struct is an opaque data structure to represent a
condition. A #GCond is an object, that threads can block on, if they
find a certain condition to be false. If other threads change the
state of this condition they can signal the #GCond, such that the
waiting thread is woken up. 
</para>

<para>
<example>
<title>Using GCond to block a thread until a condition is satisfied</title>
<programlisting>
GCond* data_cond = NULL;   /* Must be initialized somewhere */
GMutex* data_mutex = NULL; /* Must be initialized somewhere */
gpointer current_data = NULL;

void push_data (gpointer data)
{
  g_mutex_lock (data_mutex);
  current_data = data;
  g_cond_signal (data_cond);
  g_mutex_unlock (data_mutex);
}

gpointer pop_data ()
{
  gpointer data;

  g_mutex_lock (data_mutex);
  while (!current_data)
      g_cond_wait (data_cond, data_mutex);
  data = current_data;
  current_data = NULL;
  g_mutex_unlock (data_mutex);
  return data;
}
</programlisting>
</example>
</para>

<para>
Whenever a thread calls pop_data() now, it will wait until
current_data is non-NULL, i.e. until some other thread has called
push_data().
</para>

<note>
<para>
It is important to use the g_cond_wait() and g_cond_timed_wait()
functions only inside a loop, which checks for the condition to be
true as it is not guaranteed that the waiting thread will find it
fulfilled, even if the signaling thread left the condition
in that state. This is because another thread can have altered the
condition, before the waiting thread got the chance to be woken up,
even if the condition itself is protected by a #GMutex, like above.
</para>
</note>

<para>
A #GCond should only be accessed via the following functions.
</para>

<note>
<para>
All of the g_cond_* functions are actually macros. Apart from taking
the addresses of them, you can however use them as if they were functions.
</para>
</note>


<!-- ##### MACRO g_cond_new ##### -->

<para>
Creates a new #GCond. This function will abort, if g_thread_init()
has not been called yet.
</para>

@Returns: a new #GCond.


<!-- ##### MACRO g_cond_signal ##### -->
<para>
If threads are waiting for @cond, exactly one of them is woken up. It
is good practice to hold the same lock as the waiting thread, while
calling this function, though not required.
</para>

<para>
This function can also be used, if g_thread_init() has
not yet been called and will do nothing then.
</para>

@cond: a #GCond.


<!-- ##### MACRO g_cond_broadcast ##### -->

<para>
If threads are waiting for @cond, all of them are woken up. It is good
practice to lock the same mutex as the waiting threads, while calling
this function, though not required.
</para>

<para>
This function can also be used, if g_thread_init() has
not yet been called and will do nothing then.
</para>

@cond: a #GCond.


<!-- ##### MACRO g_cond_wait ##### -->

<para>
Waits until this thread is woken up on the #GCond. The #GMutex is
unlocked before falling asleep and locked again before resuming.
</para>

<para>
This function can also be used, if g_thread_init() has not yet been
called and will immediately return then.
</para>

@cond: a #GCond.
@mutex: the #GMutex, that is currently locked.


<!-- ##### MACRO g_cond_timed_wait ##### -->

<para>
Waits until this thread is woken up on the #GCond, but not longer than
until the time, that is specified by @abs_time. The #GMutex is
unlocked before falling asleep and locked again before resuming.
</para>

<para>
If @abs_time is NULL, g_cond_timed_wait() acts like g_cond_wait().
</para>

<para>
This function can also be used, if g_thread_init() has not yet been
called and will immediately return TRUE then.
</para>

@cond: a #GCond.
@mutex: the #GMutex, that is currently locked.
@abs_time: a #GTimeVal, determining the final time.
@Returns: TRUE, if the thread is woken up in time.


<!-- ##### MACRO g_cond_free ##### -->

<para>
Destroys the #GCond.
</para>

@cond: a #GCond.


<!-- ##### STRUCT GPrivate ##### -->
<para>
The #GPrivate struct is an opaque data structure to represent a thread
private data key. Threads can thereby obtain and set a pointer, which
is private to the current thread. Take our give_me_next_number()
example from above.  Now we don't want current_number to be shared
between the threads, but to be private to each thread. This can be
done as follows:

<para>
<example>
<title>Using GPrivate for per-thread data</title>
<programlisting>
  GPrivate* current_number_key = NULL; /* Must be initialized somewhere */
                                       /* with g_private_new (g_free); */

  int give_me_next_number ()
  {
    int *current_number = g_private_get (current_number_key);

    if (!current_number)
    {
      current_number = g_new (int,1);
      *current_number = 0;
      g_private_set (current_number_key, current_number);
    }
    *current_number = calc_next_number (*current_number); 
    return *current_number;
  }
</programlisting>
</example>
</para>

<para>
Here the pointer belonging to the key current_number_key is read. If
it is NULL, it has not been set yet. Then get memory for an integer
value, assign this memory to the pointer and write the pointer
back. Now we have an integer value, that is private to the current
thread.
</para>

<para>
The #GPrivate struct should only be accessed via the following functions.
</para>

<note>
<para>
All of the g_private_* functions are actually macros. Apart from taking
the addresses of them, you can however use them as if they were functions.
</para>
</note>


<!-- ##### MACRO g_private_new ##### -->

<para>
Creates a new #GPrivate. If @destructor is non-NULL, it is a pointer
to a destructor function. Whenever a thread ends and the corresponding
pointer keyed to this instance of #GPrivate is non-NULL, the
destructor is called with this pointer as the argument.
</para>

<note>
<para>
The @destructor is working quite differently from @notify in
g_static_private_set().
</para>
</note>

<note>
<para>
A #GPrivate can not be destroyed. Reuse it instead, if you can to
avoid shortage.
</para>
</note>

<note>
<para>
This function will abort, if g_thread_init() has not been called yet.
</para>
</note>

@destructor: a function to handle the data keyed to #GPrivate, when a
thread ends.


<!-- ##### MACRO g_private_get ##### -->

<para>
Returns the pointer keyed to @private_key for the current thread. This
pointer is NULL, when g_private_set() hasn't been called for the
current @private_key and thread yet.
</para>

<para>
This function can also be used, if g_thread_init() has not yet been
called and will return the value of @private_key casted to #gpointer then.
</para>

@private_key: a #GPrivate.
@Returns: the corresponding pointer.


<!-- ##### MACRO g_private_set ##### -->

<para>
Sets the pointer keyed to @private_key for the current thread.
</para>

<para>
This function can also be used, if g_thread_init() has not yet been
called and will set @private_key to @data casted to #GPrivate* then.
</para>

@private_key: a #GPrivate.
@value: 
<!-- # Unused Parameters # -->
@data: the new pointer.


<!-- ##### STRUCT GStaticPrivate ##### -->

<para>
A #GStaticPrivate works almost like a #GPrivate, but it has one
significant advantage. It doesn't need to be created at run-time like
a #GPrivate, but can be defined at compile-time. This is similar to
the difference between #GMutex and #GStaticMutex. Now look at our
give_me_next_number() example with #GStaticPrivate:
</para>

<para>
<example>
<title>Using GStaticPrivate for per-thread data</title>
<programlisting>
  int give_me_next_number ()
  {
    static GStaticPrivate current_number_key = G_STATIC_PRIVATE_INIT;
    int *current_number = g_static_private_get (&amp;current_number_key);

    if (!current_number)
    {
      current_number = g_new (int,1);
      *current_number = 0;
      g_static_private_set (&amp;current_number_key, current_number, g_free);
    }
    *current_number = calc_next_number (*current_number); 
    return *current_number;
  }
</programlisting>
</example>
</para>

@index: 

<!-- ##### MACRO G_STATIC_PRIVATE_INIT ##### -->
<para>
Every #GStaticPrivate must be initialized with this macro, before it can
be used.
</para>

<para>
<informalexample>
<programlisting>
GStaticPrivate my_private = G_STATIC_PRIVATE_INIT;
</programlisting>
</informalexample>
</para>



<!-- ##### FUNCTION g_static_private_get ##### -->
<para>
Works like g_private_get() only for a #GStaticPrivate.
</para>

<para>
This function also works, if g_thread_init() has not yet been called.
</para>

@private_key: a #GStaticPrivate.
@Returns: the corresponding pointer.


<!-- ##### FUNCTION g_static_private_set ##### -->
<para>
Sets the pointer keyed to @private_key for the current thread and the
function @notify to be called with that pointer (NULL or non-NULL),
whenever the pointer is set again or whenever the current thread ends.
</para>

<para>
This function also works, if g_thread_init() has not yet been
called. If g_thread_init() is called later, the @data keyed to
@private_key will be inherited only by the main thread, i.e. the one that
called g_thread_init().
</para>

<note>
<para>
The @notify is working quite differently from @destructor in
g_private_new().
</para>
</note>

@private_key: a #GStaticPrivate.
@data: the new pointer.
@notify: a function to be called with the pointer, whenever the
current thread ends or sets this pointer again.