<page xmlns="http://projectmallard.org/1.0/"
xmlns:its="http://www.w3.org/2005/11/its"
xmlns:xi="http://www.w3.org/2003/XInclude"
type="topic"
id="threading">
<info>
<link type="guide" xref="index#specific-how-tos"/>
<credit type="author copyright">
<name>Philip Withnall</name>
<email its:translate="no">philip.withnall@collabora.co.uk</email>
<years>2015</years>
</credit>
<include href="cc-by-sa-3-0.xml" xmlns="http://www.w3.org/2001/XInclude"/>
<desc>Moving computation out of the main thread into worker threads</desc>
</info>
<title>Threading</title>
<synopsis>
<title>Summary</title>
<list>
<item><p>
Do not use threads if at all possible.
(<link xref="#when-to-use-threading"/>)
</p></item>
<item><p>
If threads have to be used, use <code>GTask</code> or
<code>GThreadPool</code> and isolate the threaded code as much as
possible.
(<link xref="#using-threading"/>)
</p></item>
<item><p>
Use <code>g_thread_join()</code> to avoid leaking thread resources if
using <code>GThread</code> manually.
(<link xref="#using-threading"/>)
</p></item>
<item><p>
Be careful about the <code>GMainContext</code> which code is executed in
if using threads. Executing code in the wrong context can cause race
conditions, or block the main loop.
(<link xref="#using-threading"/>)
</p></item>
</list>
</synopsis>
<section id="when-to-use-threading">
<title>When to Use Threading</title>
<p>
When writing projects using GLib, the default approach should be to
<em style="strong">never use threads</em>. Instead, make proper use of the
<link xref="main-contexts">GLib main context</link> which, through the use
of asynchronous operations,
allows most blocking I/O operations to continue in the background while
the main context continues to process other events. Analysis, review and
debugging of threaded code becomes very hard, very quickly.
</p>
<p>
Threading should only be necessary when using an external library which
has blocking functions which need to be called from GLib code. If the
library provides a non-blocking alternative, or one which integrates with
a
<link href="http://pubs.opengroup.org/onlinepubs/009695399/functions/poll.html"><code>poll()</code></link>
loop, that should be used in preference. If the blocking function really
must be used, a thin wrapper should be written for it to convert it to the
normal
<link href="https://developer.gnome.org/gio/stable/GAsyncResult.html"><code>GAsyncResult</code>
style</link> of GLib asynchronous function, running the blocking operation
in a worker thread.
</p>
<example>
<p>
For example:
</p>
<code mime="text/x-csrc">
int
some_blocking_function (void *param1,
void *param2);
</code>
<p>
Should be wrapped as:
</p>
<code mime="text/x-csrc">
void
some_blocking_function_async (void *param1,
void *param2,
GCancellable *cancellable,
GAsyncReadyCallback callback,
gpointer user_data);
int
some_blocking_function_finish (GAsyncResult *result,
GError **error);
</code>
<p>
With an implementation something like:
</p>
<code mime="text/x-csrc">/* Closure for the call’s parameters. */
typedef struct {
void *param1;
void *param2;
} SomeBlockingFunctionData;
static void
some_blocking_function_data_free (SomeBlockingFunctionData *data)
{
free_param (data->param1);
free_param (data->param2);
g_free (data);
}
static void
some_blocking_function_thread_cb (GTask *task,
gpointer source_object,
gpointer task_data,
GCancellable *cancellable)
{
SomeBlockingFunctionData *data = task_data;
int retval;
/* Handle cancellation. */
if (g_task_return_error_if_cancelled (task))
{
return;
}
/* Run the blocking function. */
retval = some_blocking_function (data->param1, data->param2);
g_task_return_int (task, retval);
}
void
some_blocking_function_async (void *param1,
void *param2,
GCancellable *cancellable,
GAsyncReadyCallback callback,
gpointer user_data)
{
GTask *task = NULL; /* owned */
SomeBlockingFunctionData *data = NULL; /* owned */
g_return_if_fail (validate_param (param1));
g_return_if_fail (validate_param (param2));
g_return_if_fail (cancellable == NULL || G_IS_CANCELLABLE (cancellable));
task = g_task_new (NULL, cancellable, callback, user_data);
g_task_set_source_tag (task, some_blocking_function_async);
/* Cancellation should be handled manually using mechanisms specific to
* some_blocking_function(). */
g_task_set_return_on_cancel (task, FALSE);
/* Set up a closure containing the call’s parameters. Copy them to avoid
* locking issues between the calling thread and the worker thread. */
data = g_new0 (SomeBlockingFunctionData, 1);
data->param1 = copy_param (param1);
data->param2 = copy_param (param2);
g_task_set_task_data (task, data, some_blocking_function_data_free);
/* Run the task in a worker thread and return immediately while that continues
* in the background. When it’s done it will call @callback in the current
* thread default main context. */
g_task_run_in_thread (task, some_blocking_function_thread_cb);
g_object_unref (task);
}
int
some_blocking_function_finish (GAsyncResult *result,
GError **error)
{
g_return_val_if_fail (g_task_is_valid (result,
some_blocking_function_async), -1);
g_return_val_if_fail (error == NULL || *error == NULL, -1);
return g_task_propagate_int (G_TASK (result), error);
}</code>
<p>
See the
<link href="https://developer.gnome.org/gio/stable/GAsyncResult.html"><code>GAsyncResult</code>
documentation</link> for more details. A simple way to implement the
worker thread is to use
<link href="https://developer.gnome.org/gio/stable/GTask.html"><code>GTask</code></link>
and <link href="https://developer.gnome.org/gio/stable/GTask.html#g-task-run-in-thread"><code>g_task_run_in_thread()</code></link>.
(See also: <link xref="main-contexts#gtask"/>.)
</p>
</example>
</section>
<section id="using-threading">
<title>Using Threading</title>
<p>
If <code>GTask</code> is not suitable for writing the worker thread, a
more low-level approach must be used. This should be considered very
carefully, as it is very easy to get threading code wrong in ways which
will unpredictably cause bugs at runtime, cause deadlocks, or consume too
many resources and terminate the program.
</p>
<p>
A full manual on writing threaded code is beyond the scope of this
document, but here are a number of guidelines to follow which should
reduce the potential for bugs in threading code. The overriding principle
is to reduce the amount of code and data which can be affected by
threading — for example, reducing the number of threads, the complexity of
worker thread implementation, and the amount of data shared between
threads.
</p>
<list>
<item>
<p>
Use <link href="https://developer.gnome.org/glib/stable/glib-Thread-Pools.html"><code>GThreadPool</code></link>
instead of manually creating
<link href="https://developer.gnome.org/glib/stable/glib-Threads.html"><code>GThread</code>s</link>
if possible. <code>GThreadPool</code> supports a work queue, limits on
the number of spawned threads, and automatically joins finished
threads so they are not leaked.
</p>
</item>
<item>
<p>
If it is not possible to use a <code>GThreadPool</code> (which is
rarely the case):
</p>
<list>
<item>
<p>
Use <link href="https://developer.gnome.org/glib/stable/glib-Threads.html#g-thread-try-new"><code>g_thread_try_new()</code></link>
to spawn threads, instead of
<link href="https://developer.gnome.org/glib/stable/glib-Threads.html#g-thread-new"><code>g_thread_new()</code></link>,
so errors due to the system running out of threads can be handled
gracefully rather than unconditionally aborting the program.
</p>
</item>
<item>
<p>
Explicitly join threads using
<link href="https://developer.gnome.org/glib/stable/glib-Threads.html#g-thread-join"><code>g_thread_join()</code></link>
to avoid leaking the thread resources.
</p>
</item>
</list>
</item>
<item>
<p>
Use message passing to transfer data between threads, rather than
manual locking with mutexes. <code>GThreadPool</code> explicitly
supports this with
<link href="https://developer.gnome.org/glib/stable/glib-Thread-Pools.html#g-thread-pool-push"><code>g_thread_pool_push()</code></link>.
</p>
</item>
<item>
<p>
If mutexes must be used:
</p>
<list>
<item>
<p>
Isolate threading code as much as possible, keeping mutexes
private within classes, and tightly bound to very specific class
members.
</p>
</item>
<item>
<p>
All mutexes should be clearly commented beside their declaration,
indicating which other structures or variables they protect access
to. Similarly, those variables should be commented saying that
they should <em>only</em> be accessed with that mutex held.
</p>
</item>
</list>
</item>
<item>
<p>
Be careful about interactions between main contexts and threads. For
example,
<link href="https://developer.gnome.org/glib/stable/glib-The-Main-Event-Loop.html#g-timeout-add-seconds"><code>g_timeout_add_seconds()</code></link>
adds a timeout <em>to be executed in the global default main
context</em>, which is being run in the main thread, <em>not
necessarily</em> the current thread. Getting this wrong can mean that
work intended for a worker thread accidentally ends up being executed
in the main thread anyway. (See also:
<link xref="main-contexts#default-contexts"/>.)
</p>
</item>
</list>
</section>
<section id="debugging">
<title>Debugging</title>
<p>
Debugging threading issues is tricky, both because they are hard to
reproduce, and because they are hard to reason about. This is one of the
big reasons for avoiding using threads in the first place.
</p>
<p>
However, if a threading issue does arise,
<link xref="tooling#helgrind-and-drd">Valgrind’s drd and helgrind tools
are useful</link>.
</p>
</section>
</page>