<?xml version="1.0"?>
<!-- vim: set ts=2 sw=2 expandtab: --> 
<!DOCTYPE refentry PUBLIC "-//OASIS//DTD DocBook XML V4.5//EN"
               "http://www.oasis-open.org/docbook/xml/4.5/docbookx.dtd" [
]>
<refentry id="using-threads">
  <refmeta>
    <refentrytitle>VIPS and threads</refentrytitle>
    <manvolnum>3</manvolnum>
    <refmiscinfo>VIPS Library</refmiscinfo>
  </refmeta>

  <refnamediv>
    <refname>Using VIPS</refname>
    <refpurpose>VIPS and threading</refpurpose>
  </refnamediv>

  <refsect3 id="using-threads-intro">
    <title>Introduction</title>
    <para>
      This section tries to summarise the rules for threaded programs using 
      libvips. Generally, libvips is threaded and thread-safe, with a few 
      exceptions.
    </para>
  </refsect3>

  <refsect3 id="using-threads-images">
    <title>Images</title>
    <para>
      On startup, you need to call VIPS_INIT() single-threaded. After that,
      you can freely create images in any thread and read them in any other 
      thread. See the example at the end of this chapter. 
      Note that results can also be shared between threads for you by the vips 
      operation cache. 
    </para>

    <para>
      The exception is the drawing operators, such as vips_draw_circle(). 
      These operations modify their image argument so you can't call them on 
      the same image from more than one thread. Reading from an image while 
      another thread is writing to it with one of the draw operations will 
      obviously also fail.
    </para>

    <para>
      When libvips calculates an image, by default it will use as many 
      threads as you have CPU cores. Use vips_concurrency_set() to change this.
    </para>

  </refsect3>

  <refsect3 id="using-threads-errors">
    <title>Error handling</title>
    <para>
      libvips has a single error code (-1 or %NULL) returned by all functions 
      on error. Error messages are not returned, instead they are logged  
      in a single global error buffer shared by all threads, see 
      vips_error_buffer(). 
    </para>

    <para>
      This makes error handling very simple but the obvious downside is that
      because error returns and error messages are separate when you 
      detect an error return you can't be 
      sure that what's in the error buffer is the message that matches your
      error. 
    </para>

    <para>
      The simplest way to handle this is to present the whole error log to 
      the user on the next interaction and leave it to them to decide what 
      action caused the failure. 
    </para>

  </refsect3>

  <refsect3 id="using-threads-regions">
    <title>Using #VipsRegion between threads</title>
    <para>
      #VipsImage objects are immutable and can be shared between 
      threads very simply.
      However the lower-level #VipsRegion object used to implement #VipsImage 
      (see <link linkend="extending">Extending VIPS</link>) is mutable and you 
      can only use a #VipsRegion from one thread at once. 
    </para>

    <para>
      In fact it's worse than that: to reduce locking, #VipsRegion keeps a 
      lot of state in per-thread storage. If you want to create a region in 
      one thread and use it in another, you have to first tag the region as 
      unowned from the creating thread with vips__region_no_ownership(), then 
      in the recieving thread take ownership with 
      vips__region_take_ownership(). See the source for operations like 
      vips_tilecache() if you're curious how this works. 
    </para>

    <para>
      libvips includes a set of sanity checks for region ownership and will
      fail if you don't pass ownership correctly. 
    </para>

  </refsect3>

  <refsect3 id="using-threads-example">
    <title>Example</title>
    <para>
      This example runs many vips_resize() in parallel from many threads. 
    </para>

<example>
<title>VIPS and threads example</title>
<programlisting language="C">
/* Read from many threads. 
 *
 * Compile with:
 *
 * 	gcc -g -Wall soak.c `pkg-config vips --cflags --libs`
 *
 * Run with:
 *
 * 	rm -rf x
 * 	mkdir x
 * 	for i in {0..10}; do ./a.out ~/pics/k2.jpg; done
 *
 */

#include &lt;stdio.h&gt;
#include &lt;glib.h&gt;

#include &lt;vips/vips.h&gt;

/* How many pings we run at once.
 */
#define NUM_IN_PARALLEL (50)

/* Number of tests we do in total.
 */
#define TOTAL_TESTS (NUM_IN_PARALLEL * 20)

/* Workers queue up on this.
 */
GMutex allocation_lock;

/* Our set of threads.
 */
GThread *workers[NUM_IN_PARALLEL];

/* Number of calls so far.
 */
int n_calls = 0;

/* Our test function. This is called by NUM_IN_PARALLEL threads a total of
 * TOTAL_TESTS times. 
 */
static int
test (const char *filename)
{
  VipsImage *im, *x;
  char output_file[256];

  snprintf (output_file, 256, "x/tmp-%p.jpg", g_thread_self ());

  if (!(im = vips_image_new_from_file (filename,
				       "access",
				       VIPS_ACCESS_SEQUENTIAL_UNBUFFERED,
				       NULL)))
    return (-1);

  if (vips_resize (im, &amp;x, 0.1, NULL))
    {
      g_object_unref (im);
      return (-1);
    }
  g_object_unref (im);
  im = x;

  if (vips_image_write_to_file (im, output_file, NULL))
    {
      g_object_unref (im);
      return (-1);
    }
  g_object_unref (im);

  return (0);
}

/* What we run as a thread.
 */
static void *
worker (void *data)
{
  const char *filename = (const char *) data;

  for (;;)
    {
      gboolean done;

      done = FALSE;
      g_mutex_lock (&amp;allocation_lock);
      n_calls += 1;
      if (n_calls &gt; TOTAL_TESTS)
	done = TRUE;
      g_mutex_unlock (&amp;allocation_lock);

      if (done)
	break;

      if (test (filename))
	vips_error_exit (NULL);
    }

  return (NULL);
}

int
main (int argc, char **argv)
{
  int i;

  if (VIPS_INIT (argv[0]))
    vips_error_exit (NULL);

  g_mutex_init (&amp;allocation_lock);

  for (i = 0; i &lt; NUM_IN_PARALLEL; i++)
    workers[i] = g_thread_new (NULL, (GThreadFunc) worker, argv[1]);

  for (i = 0; i &lt; NUM_IN_PARALLEL; i++)
    g_thread_join (workers[i]);

  return (0);
}
</programlisting>
</example>

  </refsect3>

</refentry>