gstreamer/gst/tcp/gstmultifdsink.c

/* GStreamer
 * Copyright (C) <1999> Erik Walthinsen <omega@cse.ogi.edu>
 * Copyright (C) <2004> Thomas Vander Stichele <thomas at apestaart dot org>
 * Copyright (C) 2006 Wim Taymans <wim at fluendo dot com>
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Library General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Library General Public License for more details.
 *
 * You should have received a copy of the GNU Library General Public
 * License along with this library; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 02111-1307, USA.
 */

/**
 * SECTION:element-multifdsink
 * @short_description: Send data to multiple file descriptors
 * @see_also: tcpserversink
 *
 * <refsect2>
 * <para>
 * This plugin writes incoming data to a set of file descriptors. The
 * file descriptors can be added to multifdsink by emiting the "add" signal. For
 * each descriptor added, the "client-added" signal will be called.
 * </para>
 * <para>
 * As of version 0.10.8, a client can also be added with the "add-full" signal
 * that allows for more control over what and how much data a client 
 * initially receives.
 * </para>
 * <para>
 * Clients can be removed from multifdsink by emiting the "remove" signal. For
 * each descriptor removed, the "client-removed" signal will be called. The
 * "client-removed" signal can also be fired when multifdsink decides that a
 * client is not active anymore or, depending on the value of the
 * "recover-policy," if the client is reading to slow.
 * In all cases, multifdsink will never ever close a file descriptor itself.
 * The user of multifdsink is responsible for closing all file descriptors.
 * This can for example be done in response to the "client-fd-removed" signal.
 * Note that multifdsink still has a reference to the file descriptor when the
 * "client-removed" signal is emited so that "get-stats" can be performed on
 * the descriptor; It is therefore not allowed to close the file descriptor in
 * the "client-removed" signal, use the "client-fd-removed" signal to safely 
 * close the fd.
 * </para>
 * <para>
 * Multifdsink internally keeps a queue of the incomming buffers and uses a
 * separate thread to send the buffers to the clients. This ensures that no
 * client write can block the pipeline and that clients can read with different
 * speeds.
 * </para>
 * <para>
 * When adding a client to multifdsink, the "sync-method" property will define
 * which buffer in the queued buffers will be sent first to the client. Clients 
 * can be sent respectively the most recent buffer (which might not be decodable
 * by the client when it is not a keyframe), the next keyframe received in 
 * multifdsink (which can take some time depending on the keyframe rate, or the
 * last received keyframe (which will cause a simpl burst-on-connect). 
 * Multifdsink will always keep at least one keyframe in its internal buffers
 * when the sync-mode is set to latest-keyframe.
 * </para>
 * <para>
 * Multifdsink can be instructed to keep at least a minimum amount of data
 * expressed in time or byte units in its internal queues with the the 
 * "time-min" and "bytes-min" properties respectively. These properties are
 * usefull if the application adds clients with the "add-full" signal to
 * make sure that a burst connect can actually be honored. 
 * </para>
 * <para>
 * When streaming data, clients are allowed to read at a different rate than
 * the rate at which multifdsink receives data. If the client is reading too
 * fast, no data will be send to the client until multifdsink receives more
 * data. If the client however reads too slow, data for that client will bunch
 * up in multifdsink. Two properties control the amount of data (buffers) that
 * is queued in multifdsink: "buffers-max" and "buffers-soft-max". A client
 * with a lag of "buffers-max" is removed from multifdsink forcibly.
 * </para>
 * <para>
 * A client with a lag of at least "buffers-soft-max" enters the recovery
 * procedure which is controlled with the "recover-policy" property. A recover
 * policy of NONE will do nothing, RESYNC_LATEST will send the most recently
 * received buffer as the next buffer for the client, RESYNC_SOFT_LIMIT
 * positions the client to the soft limit in the buffer queue and
 * RESYNC_KEYFRAME positions the client to the most recent keyframe in the
 * buffer queue.
 * </para>
 * <para>
 * multifdsink will by default synchronize on the clock before serving the buffers
 * to the clients. This behaviour can be disabled by setting the sync 
 * property to FALSE. Multifdsink will be default not do QoS and will never
 * drop late buffers.
 * </para>
 * </refsect2>
 *
 * Last reviewed on 2006-06-13 (0.10.9)
 */

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <gst/gst-i18n-plugin.h>

#include <sys/ioctl.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <sys/stat.h>

#ifdef HAVE_FIONREAD_IN_SYS_FILIO
#include <sys/filio.h>
#endif

#include "gstmultifdsink.h"
#include "gsttcp-marshal.h"

#define NOT_IMPLEMENTED 0

/* the select call is also performed on the control sockets, that way
 * we can send special commands to unblock or restart the select call */
#define CONTROL_RESTART         'R'     /* restart the select call */
#define CONTROL_STOP            'S'     /* stop the select call */
#define CONTROL_SOCKETS(sink)   sink->control_sock
#define WRITE_SOCKET(sink)      sink->control_sock[1]
#define READ_SOCKET(sink)       sink->control_sock[0]

#define SEND_COMMAND(sink, command)             \
G_STMT_START {                                  \
  unsigned char c; c = command;                 \
  write (WRITE_SOCKET(sink).fd, &c, 1);         \
} G_STMT_END

#define READ_COMMAND(sink, command, res)        \
G_STMT_START {                                  \
  res = read(READ_SOCKET(sink).fd, &command, 1);\
} G_STMT_END

/* elementfactory information */
static const GstElementDetails gst_multi_fd_sink_details =
GST_ELEMENT_DETAILS ("Multi filedescriptor sink",
    "Sink/Network",
    "Send data to multiple filedescriptors",
    "Thomas Vander Stichele <thomas at apestaart dot org>, "
    "Wim Taymans <wim@fluendo.com>");

static GstStaticPadTemplate sinktemplate = GST_STATIC_PAD_TEMPLATE ("sink",
    GST_PAD_SINK,
    GST_PAD_ALWAYS,
    GST_STATIC_CAPS_ANY);

GST_DEBUG_CATEGORY (multifdsink_debug);
#define GST_CAT_DEFAULT (multifdsink_debug)

/* MultiFdSink signals and args */
enum
{
  /* methods */
  SIGNAL_ADD,
  SIGNAL_ADD_BURST,
  SIGNAL_REMOVE,
  SIGNAL_CLEAR,
  SIGNAL_GET_STATS,

  /* signals */
  SIGNAL_CLIENT_ADDED,
  SIGNAL_CLIENT_REMOVED,
  SIGNAL_CLIENT_FD_REMOVED,

  LAST_SIGNAL
};

/* this is really arbitrarily chosen */
#define DEFAULT_PROTOCOL                GST_TCP_PROTOCOL_NONE
#define DEFAULT_MODE                    GST_FDSET_MODE_POLL
#define DEFAULT_BUFFERS_MAX             -1
#define DEFAULT_BUFFERS_SOFT_MAX        -1
#define DEFAULT_TIME_MIN                -1
#define DEFAULT_BYTES_MIN               -1
#define DEFAULT_BUFFERS_MIN             -1
#define DEFAULT_UNIT_TYPE               GST_UNIT_TYPE_BUFFERS
#define DEFAULT_UNITS_MAX               -1
#define DEFAULT_UNITS_SOFT_MAX          -1
#define DEFAULT_RECOVER_POLICY          GST_RECOVER_POLICY_NONE
#define DEFAULT_TIMEOUT                 0
#define DEFAULT_SYNC_METHOD             GST_SYNC_METHOD_LATEST

#define DEFAULT_BURST_UNIT              GST_UNIT_TYPE_UNDEFINED
#define DEFAULT_BURST_VALUE             0

enum
{
  PROP_0,
  PROP_PROTOCOL,
  PROP_MODE,
  PROP_BUFFERS_QUEUED,
  PROP_BYTES_QUEUED,
  PROP_TIME_QUEUED,

  PROP_UNIT_TYPE,
  PROP_UNITS_MAX,
  PROP_UNITS_SOFT_MAX,

  PROP_BUFFERS_MAX,
  PROP_BUFFERS_SOFT_MAX,

  PROP_TIME_MIN,
  PROP_BYTES_MIN,
  PROP_BUFFERS_MIN,

  PROP_RECOVER_POLICY,
  PROP_TIMEOUT,
  PROP_SYNC_METHOD,
  PROP_BYTES_TO_SERVE,
  PROP_BYTES_SERVED,

  PROP_BURST_UNIT,
  PROP_BURST_VALUE,
};

#define GST_TYPE_RECOVER_POLICY (gst_recover_policy_get_type())
static GType
gst_recover_policy_get_type (void)
{
  static GType recover_policy_type = 0;
  static const GEnumValue recover_policy[] = {
    {GST_RECOVER_POLICY_NONE,
        "Do not try to recover", "none"},
    {GST_RECOVER_POLICY_RESYNC_LATEST,
        "Resync client to latest buffer", "latest"},
    {GST_RECOVER_POLICY_RESYNC_SOFT_LIMIT,
        "Resync client to soft limit", "soft-limit"},
    {GST_RECOVER_POLICY_RESYNC_KEYFRAME,
        "Resync client to most recent keyframe", "keyframe"},
    {0, NULL, NULL},
  };

  if (!recover_policy_type) {
    recover_policy_type =
        g_enum_register_static ("GstRecoverPolicy", recover_policy);
  }
  return recover_policy_type;
}

#define GST_TYPE_SYNC_METHOD (gst_sync_method_get_type())
static GType
gst_sync_method_get_type (void)
{
  static GType sync_method_type = 0;
  static const GEnumValue sync_method[] = {
    {GST_SYNC_METHOD_LATEST,
        "Serve starting from the latest buffer", "latest"},
    {GST_SYNC_METHOD_NEXT_KEYFRAME,
        "Serve starting from the next keyframe", "next-keyframe"},
    {GST_SYNC_METHOD_LATEST_KEYFRAME,
          "Serve everything since the latest keyframe (burst)",
        "latest-keyframe"},
    {GST_SYNC_METHOD_BURST, "Serve burst-value data to client", "burst"},
    {GST_SYNC_METHOD_BURST_KEYFRAME,
          "Serve burst-value data starting on a keyframe",
        "burst-keyframe"},
    {GST_SYNC_METHOD_BURST_WITH_KEYFRAME,
          "Serve burst-value data preferably starting on a keyframe",
        "burst-with-keyframe"},
    {0, NULL, NULL},
  };

  if (!sync_method_type) {
    sync_method_type = g_enum_register_static ("GstSyncMethod", sync_method);
  }
  return sync_method_type;
}

#define GST_TYPE_UNIT_TYPE (gst_unit_type_get_type())
static GType
gst_unit_type_get_type (void)
{
  static GType unit_type_type = 0;
  static const GEnumValue unit_type[] = {
    {GST_UNIT_TYPE_UNDEFINED, "Undefined", "undefined"},
    {GST_UNIT_TYPE_BUFFERS, "Buffers", "buffers"},
    {GST_UNIT_TYPE_BYTES, "Bytes", "bytes"},
    {GST_UNIT_TYPE_TIME, "Time", "time"},
    {0, NULL, NULL},
  };

  if (!unit_type_type) {
    unit_type_type = g_enum_register_static ("GstTCPUnitType", unit_type);
  }
  return unit_type_type;
}

#define GST_TYPE_CLIENT_STATUS (gst_client_status_get_type())
static GType
gst_client_status_get_type (void)
{
  static GType client_status_type = 0;
  static const GEnumValue client_status[] = {
    {GST_CLIENT_STATUS_OK, "ok", "ok"},
    {GST_CLIENT_STATUS_CLOSED, "Closed", "closed"},
    {GST_CLIENT_STATUS_REMOVED, "Removed", "removed"},
    {GST_CLIENT_STATUS_SLOW, "Too slow", "slow"},
    {GST_CLIENT_STATUS_ERROR, "Error", "error"},
    {GST_CLIENT_STATUS_DUPLICATE, "Duplicate", "duplicate"},
    {0, NULL, NULL},
  };

  if (!client_status_type) {
    client_status_type =
        g_enum_register_static ("GstClientStatus", client_status);
  }
  return client_status_type;
}

static void gst_multi_fd_sink_finalize (GObject * object);

static void gst_multi_fd_sink_remove_client_link (GstMultiFdSink * sink,
    GList * link);

static GstFlowReturn gst_multi_fd_sink_render (GstBaseSink * bsink,
    GstBuffer * buf);
static GstStateChangeReturn gst_multi_fd_sink_change_state (GstElement *
    element, GstStateChange transition);

static void gst_multi_fd_sink_set_property (GObject * object, guint prop_id,
    const GValue * value, GParamSpec * pspec);
static void gst_multi_fd_sink_get_property (GObject * object, guint prop_id,
    GValue * value, GParamSpec * pspec);

GST_BOILERPLATE (GstMultiFdSink, gst_multi_fd_sink, GstBaseSink,
    GST_TYPE_BASE_SINK);

static guint gst_multi_fd_sink_signals[LAST_SIGNAL] = { 0 };

static void
gst_multi_fd_sink_base_init (gpointer g_class)
{
  GstElementClass *element_class = GST_ELEMENT_CLASS (g_class);

  gst_element_class_add_pad_template (element_class,
      gst_static_pad_template_get (&sinktemplate));

  gst_element_class_set_details (element_class, &gst_multi_fd_sink_details);
}

static void
gst_multi_fd_sink_class_init (GstMultiFdSinkClass * klass)
{
  GObjectClass *gobject_class;
  GstElementClass *gstelement_class;
  GstBaseSinkClass *gstbasesink_class;

  gobject_class = (GObjectClass *) klass;
  gstelement_class = (GstElementClass *) klass;
  gstbasesink_class = (GstBaseSinkClass *) klass;

  gobject_class->set_property = gst_multi_fd_sink_set_property;
  gobject_class->get_property = gst_multi_fd_sink_get_property;
  gobject_class->finalize = gst_multi_fd_sink_finalize;

  g_object_class_install_property (gobject_class, PROP_PROTOCOL,
      g_param_spec_enum ("protocol", "Protocol", "The protocol to wrap data in",
          GST_TYPE_TCP_PROTOCOL, DEFAULT_PROTOCOL, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_MODE,
      g_param_spec_enum ("mode", "Mode",
          "The mode for selecting activity on the fds", GST_TYPE_FDSET_MODE,
          DEFAULT_MODE, G_PARAM_READWRITE));

  g_object_class_install_property (gobject_class, PROP_BUFFERS_MAX,
      g_param_spec_int ("buffers-max", "Buffers max",
          "max number of buffers to queue for a client (-1 = no limit)", -1,
          G_MAXINT, DEFAULT_BUFFERS_MAX, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class,
      PROP_BUFFERS_SOFT_MAX, g_param_spec_int ("buffers-soft-max",
          "Buffers soft max",
          "Recover client when going over this limit (-1 = no limit)", -1,
          G_MAXINT, DEFAULT_BUFFERS_SOFT_MAX, G_PARAM_READWRITE));

  g_object_class_install_property (gobject_class, PROP_BYTES_MIN,
      g_param_spec_int ("bytes-min", "Bytes min",
          "min number of bytes to queue (-1 = as little as possible)", -1,
          G_MAXINT, DEFAULT_BYTES_MIN, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_TIME_MIN,
      g_param_spec_int64 ("time-min", "Time min",
          "min number of time to queue (-1 = as little as possible)", -1,
          G_MAXINT64, DEFAULT_TIME_MIN, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_BUFFERS_MIN,
      g_param_spec_int ("buffers-min", "Buffers min",
          "min number of buffers to queue (-1 = as few as possible)", -1,
          G_MAXINT, DEFAULT_BUFFERS_MIN, G_PARAM_READWRITE));

#if NOT_IMPLEMENTED
  g_object_class_install_property (gobject_class, PROP_UNIT_TYPE,
      g_param_spec_enum ("unit-type", "Units type",
          "The unit to measure the max/soft-max/queued properties",
          GST_TYPE_UNIT_TYPE, DEFAULT_UNIT_TYPE, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_UNITS_MAX,
      g_param_spec_int ("units-max", "Units max",
          "max number of units to queue (-1 = no limit)", -1, G_MAXINT,
          DEFAULT_UNITS_MAX, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_UNITS_SOFT_MAX,
      g_param_spec_int ("units-soft-max", "Units soft max",
          "Recover client when going over this limit (-1 = no limit)", -1,
          G_MAXINT, DEFAULT_UNITS_SOFT_MAX, G_PARAM_READWRITE));
#endif

  g_object_class_install_property (gobject_class, PROP_BUFFERS_QUEUED,
      g_param_spec_uint ("buffers-queued", "Buffers queued",
          "Number of buffers currently queued", 0, G_MAXUINT, 0,
          G_PARAM_READABLE));
#if NOT_IMPLEMENTED
  g_object_class_install_property (gobject_class, PROP_BYTES_QUEUED,
      g_param_spec_uint ("bytes-queued", "Bytes queued",
          "Number of bytes currently queued", 0, G_MAXUINT, 0,
          G_PARAM_READABLE));
  g_object_class_install_property (gobject_class, PROP_TIME_QUEUED,
      g_param_spec_uint64 ("time-queued", "Time queued",
          "Number of time currently queued", 0, G_MAXUINT64, 0,
          G_PARAM_READABLE));
#endif

  g_object_class_install_property (gobject_class, PROP_RECOVER_POLICY,
      g_param_spec_enum ("recover-policy", "Recover Policy",
          "How to recover when client reaches the soft max",
          GST_TYPE_RECOVER_POLICY, DEFAULT_RECOVER_POLICY, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_TIMEOUT,
      g_param_spec_uint64 ("timeout", "Timeout",
          "Maximum inactivity timeout in nanoseconds for a client (0 = no limit)",
          0, G_MAXUINT64, DEFAULT_TIMEOUT, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_SYNC_METHOD,
      g_param_spec_enum ("sync-method", "Sync Method",
          "How to sync new clients to the stream",
          GST_TYPE_SYNC_METHOD, DEFAULT_SYNC_METHOD, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_BYTES_TO_SERVE,
      g_param_spec_uint64 ("bytes-to-serve", "Bytes to serve",
          "Number of bytes received to serve to clients", 0, G_MAXUINT64, 0,
          G_PARAM_READABLE));
  g_object_class_install_property (gobject_class, PROP_BYTES_SERVED,
      g_param_spec_uint64 ("bytes-served", "Bytes served",
          "Total number of bytes send to all clients", 0, G_MAXUINT64, 0,
          G_PARAM_READABLE));

  g_object_class_install_property (gobject_class, PROP_BURST_UNIT,
      g_param_spec_enum ("burst-unit", "Burst unit",
          "The format of the burst units (when sync-method is burst[[-with]-keyframe])",
          GST_TYPE_UNIT_TYPE, DEFAULT_BURST_UNIT, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_BURST_VALUE,
      g_param_spec_uint64 ("burst-value", "Burst value",
          "The amount of burst expressed in burst-unit",
          0, G_MAXUINT64, DEFAULT_BURST_VALUE, G_PARAM_READWRITE));

  /**
   * GstMultiFdSink::add:
   * @gstmultifdsink: the multifdsink element to emit this signal on
   * @fd:             the file descriptor to add to multifdsink
   *
   * Hand the given open file descriptor to multifdsink to write to.
   */
  gst_multi_fd_sink_signals[SIGNAL_ADD] =
      g_signal_new ("add", G_TYPE_FROM_CLASS (klass), G_SIGNAL_RUN_LAST,
      G_STRUCT_OFFSET (GstMultiFdSinkClass, add),
      NULL, NULL, g_cclosure_marshal_VOID__INT, G_TYPE_NONE, 1, G_TYPE_INT);
  /**
   * GstMultiFdSink::add-full:
   * @gstmultifdsink: the multifdsink element to emit this signal on
   * @fd:             the file descriptor to add to multifdsink
   * @keyframe:       start bursting from a keyframe
   * @unit_type:      the unit-type of @value
   * @value:          the minimal amount of data to burst expressed in
   *                  @format units.
   *
   * Hand the given open file descriptor to multifdsink to write to and
   * specify the burst parameters for the new connection.
   */
  gst_multi_fd_sink_signals[SIGNAL_ADD_BURST] =
      g_signal_new ("add-full", G_TYPE_FROM_CLASS (klass), G_SIGNAL_RUN_LAST,
      G_STRUCT_OFFSET (GstMultiFdSinkClass, add_full),
      NULL, NULL, gst_tcp_marshal_VOID__INT_BOOLEAN_INT_UINT64_INT_UINT64,
      G_TYPE_NONE, 6, G_TYPE_INT, G_TYPE_BOOLEAN, GST_TYPE_UNIT_TYPE,
      G_TYPE_UINT64, GST_TYPE_UNIT_TYPE, G_TYPE_UINT64);
  /**
   * GstMultiFdSink::remove:
   * @gstmultifdsink: the multifdsink element to emit this signal on
   * @fd:             the file descriptor to remove from multifdsink
   *
   * Remove the given open file descriptor from multifdsink.
   */
  gst_multi_fd_sink_signals[SIGNAL_REMOVE] =
      g_signal_new ("remove", G_TYPE_FROM_CLASS (klass), G_SIGNAL_RUN_LAST,
      G_STRUCT_OFFSET (GstMultiFdSinkClass, remove),
      NULL, NULL, gst_tcp_marshal_VOID__INT, G_TYPE_NONE, 1, G_TYPE_INT);
  /**
   * GstMultiFdSink::clear:
   * @gstmultifdsink: the multifdsink element to emit this signal on
   *
   * Remove all file descriptors from multifdsink.  Since multifdsink did not
   * open fd's itself, it does not explicitly close the fd.  The application
   * should do so by connecting to the client-removed callback.
   */
  gst_multi_fd_sink_signals[SIGNAL_CLEAR] =
      g_signal_new ("clear", G_TYPE_FROM_CLASS (klass), G_SIGNAL_RUN_LAST,
      G_STRUCT_OFFSET (GstMultiFdSinkClass, clear),
      NULL, NULL, g_cclosure_marshal_VOID__VOID, G_TYPE_NONE, 0);

  /**
   * GstMultiFdSink::get-stats:
   * @gstmultifdsink: the multifdsink element to emit this signal on
   * @fd:             the file descriptor to get stats of from multifdsink
   *
   * Get statistics about @fd. This function returns a GValueArray to ease
   * automatic wrapping for bindings.
   *
   * Returns: a GValueArray with the statistics. The array contains 5 guint64
   *     values that represent respectively total number of bytes sent, time
   *     when the client was added, time when the client was disconnected/removed,
   *     time the client is/was active, last activity time. All times are
   *     expressed in nanoseconds (GstClockTime).
   */
  gst_multi_fd_sink_signals[SIGNAL_GET_STATS] =
      g_signal_new ("get-stats", G_TYPE_FROM_CLASS (klass), G_SIGNAL_RUN_LAST,
      G_STRUCT_OFFSET (GstMultiFdSinkClass, get_stats),
      NULL, NULL, gst_tcp_marshal_BOXED__INT, G_TYPE_VALUE_ARRAY, 1,
      G_TYPE_INT);

  /**
   * GstMultiFdSink::client-added:
   * @gstmultifdsink: the multifdsink element that emitted this signal
   * @fd:             the file descriptor that was added to multifdsink
   *
   * The given file descriptor was added to multifdsink. This signal will
   * be emited from the streaming thread so application should be prepared
   * for that.
   */
  gst_multi_fd_sink_signals[SIGNAL_CLIENT_ADDED] =
      g_signal_new ("client-added", G_TYPE_FROM_CLASS (klass),
      G_SIGNAL_RUN_LAST, G_STRUCT_OFFSET (GstMultiFdSinkClass, client_added),
      NULL, NULL, gst_tcp_marshal_VOID__INT, G_TYPE_NONE, 1, G_TYPE_INT);
  /**
   * GstMultiFdSink::client-removed:
   * @gstmultifdsink: the multifdsink element that emitted this signal
   * @fd:             the file descriptor that is to be removed from multifdsink
   * @status:         the reason why the client was removed
   *
   * The given file descriptor is about to be removed from multifdsink. This
   * signal will be emited from the streaming thread so applications should
   * be prepared for that.
   *
   * @gstmultifdsink still holds a handle to @fd so it is possible to call
   * the get-stats signal from this callback. For the same reason it is
   * not safe to close() and reuse @fd in this callback.
   */
  gst_multi_fd_sink_signals[SIGNAL_CLIENT_REMOVED] =
      g_signal_new ("client-removed", G_TYPE_FROM_CLASS (klass),
      G_SIGNAL_RUN_LAST, G_STRUCT_OFFSET (GstMultiFdSinkClass,
          client_removed), NULL, NULL, gst_tcp_marshal_VOID__INT_BOXED,
      G_TYPE_NONE, 2, G_TYPE_INT, GST_TYPE_CLIENT_STATUS);
  /**
   * GstMultiFdSink::client-fd-removed:
   * @gstmultifdsink: the multifdsink element that emitted this signal
   * @fd:             the file descriptor that was removed from multifdsink
   *
   * The given file descriptor was removed from multifdsink. This signal will
   * be emited from the streaming thread so applications should be prepared
   * for that.
   *
   * In this callback, @gstmultifdsink has removed all the information
   * associated with @fd and it is therefore not possible to call get-stats
   * with @fd. It is however safe to close() and reuse @fd in the callback.
   *
   * Since: 0.10.7
   */
  gst_multi_fd_sink_signals[SIGNAL_CLIENT_FD_REMOVED] =
      g_signal_new ("client-fd-removed", G_TYPE_FROM_CLASS (klass),
      G_SIGNAL_RUN_LAST, G_STRUCT_OFFSET (GstMultiFdSinkClass,
          client_fd_removed), NULL, NULL, gst_tcp_marshal_VOID__INT,
      G_TYPE_NONE, 1, G_TYPE_INT);

  gstelement_class->change_state =
      GST_DEBUG_FUNCPTR (gst_multi_fd_sink_change_state);

  gstbasesink_class->render = GST_DEBUG_FUNCPTR (gst_multi_fd_sink_render);

  klass->add = GST_DEBUG_FUNCPTR (gst_multi_fd_sink_add);
  klass->add_full = GST_DEBUG_FUNCPTR (gst_multi_fd_sink_add_full);
  klass->remove = GST_DEBUG_FUNCPTR (gst_multi_fd_sink_remove);
  klass->clear = GST_DEBUG_FUNCPTR (gst_multi_fd_sink_clear);
  klass->get_stats = GST_DEBUG_FUNCPTR (gst_multi_fd_sink_get_stats);

  GST_DEBUG_CATEGORY_INIT (multifdsink_debug, "multifdsink", 0, "FD sink");
}

static void
gst_multi_fd_sink_init (GstMultiFdSink * this, GstMultiFdSinkClass * klass)
{
  GST_OBJECT_FLAG_UNSET (this, GST_MULTI_FD_SINK_OPEN);

  this->protocol = DEFAULT_PROTOCOL;
  this->mode = DEFAULT_MODE;

  CLIENTS_LOCK_INIT (this);
  this->clients = NULL;
  this->fd_hash = g_hash_table_new (g_int_hash, g_int_equal);

  this->bufqueue = g_array_new (FALSE, TRUE, sizeof (GstBuffer *));
  this->unit_type = DEFAULT_UNIT_TYPE;
  this->units_max = DEFAULT_UNITS_MAX;
  this->units_soft_max = DEFAULT_UNITS_SOFT_MAX;
  this->time_min = DEFAULT_TIME_MIN;
  this->bytes_min = DEFAULT_BYTES_MIN;
  this->buffers_min = DEFAULT_BUFFERS_MIN;
  this->recover_policy = DEFAULT_RECOVER_POLICY;

  this->timeout = DEFAULT_TIMEOUT;
  this->def_sync_method = DEFAULT_SYNC_METHOD;
  this->def_burst_unit = DEFAULT_BURST_UNIT;
  this->def_burst_value = DEFAULT_BURST_VALUE;

  this->header_flags = 0;
}

static void
gst_multi_fd_sink_finalize (GObject * object)
{
  GstMultiFdSink *this;

  this = GST_MULTI_FD_SINK (object);

  CLIENTS_LOCK_FREE (this);
  g_hash_table_destroy (this->fd_hash);
  g_array_free (this->bufqueue, TRUE);

  G_OBJECT_CLASS (parent_class)->finalize (object);
}

/* "add-full" signal implemntation */
void
gst_multi_fd_sink_add_full (GstMultiFdSink * sink, int fd,
    GstSyncMethod sync_method, GstUnitType min_unit, guint64 min_value,
    GstUnitType max_unit, guint64 max_value)
{
  GstTCPClient *client;
  GList *clink;
  GTimeVal now;
  gint flags, res;
  struct stat statbuf;

  GST_DEBUG_OBJECT (sink, "[fd %5d] adding client", fd);

  /* do limits check if we can */
  if (min_unit == max_unit) {
    if (max_value != -1 && min_value != -1 && max_value < min_value)
      goto wrong_limits;
  }

  /* create client datastructure */
  client = g_new0 (GstTCPClient, 1);
  client->fd.fd = fd;
  client->status = GST_CLIENT_STATUS_OK;
  client->bufpos = -1;
  client->bufoffset = 0;
  client->sending = NULL;
  client->bytes_sent = 0;
  client->dropped_buffers = 0;
  client->avg_queue_size = 0;
  client->new_connection = TRUE;
  client->burst_min_unit = min_unit;
  client->burst_min_value = min_value;
  client->burst_max_unit = max_unit;
  client->burst_max_value = max_value;
  client->sync_method = sync_method;

  /* update start time */
  g_get_current_time (&now);
  client->connect_time = GST_TIMEVAL_TO_TIME (now);
  client->disconnect_time = 0;
  /* set last activity time to connect time */
  client->last_activity_time = client->connect_time;

  CLIENTS_LOCK (sink);

  /* check the hash to find a duplicate fd */
  clink = g_hash_table_lookup (sink->fd_hash, &client->fd.fd);
  if (clink != NULL)
    goto duplicate;

  /* we can add the fd now */
  clink = sink->clients = g_list_prepend (sink->clients, client);
  g_hash_table_insert (sink->fd_hash, &client->fd.fd, clink);

  /* set the socket to non blocking */
  res = fcntl (fd, F_SETFL, O_NONBLOCK);
  /* we always read from a client */
  gst_fdset_add_fd (sink->fdset, &client->fd);

  /* we don't try to read from write only fds */
  flags = fcntl (fd, F_GETFL, 0);
  if ((flags & O_ACCMODE) != O_WRONLY) {
    gst_fdset_fd_ctl_read (sink->fdset, &client->fd, TRUE);
  }
  /* figure out the mode, can't use send() for non sockets */
  res = fstat (fd, &statbuf);
  if (S_ISSOCK (statbuf.st_mode)) {
    client->is_socket = TRUE;
  }

  SEND_COMMAND (sink, CONTROL_RESTART);

  CLIENTS_UNLOCK (sink);

  g_signal_emit (G_OBJECT (sink),
      gst_multi_fd_sink_signals[SIGNAL_CLIENT_ADDED], 0, fd);

  return;

  /* errors */
wrong_limits:
  {
    GST_WARNING_OBJECT (sink,
        "[fd %5d] wrong values min =%" G_GUINT64_FORMAT ", max=%"
        G_GUINT64_FORMAT ", unit %d specified when adding client", fd,
        min_value, max_value, min_unit);
    return;
  }
duplicate:
  {
    client->status = GST_CLIENT_STATUS_DUPLICATE;
    CLIENTS_UNLOCK (sink);
    GST_WARNING_OBJECT (sink, "[fd %5d] duplicate client found, refusing", fd);
    g_signal_emit (G_OBJECT (sink),
        gst_multi_fd_sink_signals[SIGNAL_CLIENT_REMOVED], 0, fd,
        client->status);
    g_free (client);
    return;
  }
}

/* "add" signal implemntation */
void
gst_multi_fd_sink_add (GstMultiFdSink * sink, int fd)
{
  gst_multi_fd_sink_add_full (sink, fd, sink->def_sync_method,
      sink->def_burst_unit, sink->def_burst_value, sink->def_burst_unit, -1);
}

/* "remove" signal implemntation */
void
gst_multi_fd_sink_remove (GstMultiFdSink * sink, int fd)
{
  GList *clink;

  GST_DEBUG_OBJECT (sink, "[fd %5d] removing client", fd);

  CLIENTS_LOCK (sink);
  clink = g_hash_table_lookup (sink->fd_hash, &fd);
  if (clink != NULL) {
    GstTCPClient *client = (GstTCPClient *) clink->data;

    client->status = GST_CLIENT_STATUS_REMOVED;
    gst_multi_fd_sink_remove_client_link (sink, clink);
    SEND_COMMAND (sink, CONTROL_RESTART);
  } else {
    GST_WARNING_OBJECT (sink, "[fd %5d] no client with this fd found!", fd);
  }
  CLIENTS_UNLOCK (sink);
}

/* can be called both through the signal (ie from any thread) or when stopping,
 * after the writing thread has shut down */
void
gst_multi_fd_sink_clear (GstMultiFdSink * sink)
{
  GList *clients, *next;

  GST_DEBUG_OBJECT (sink, "clearing all clients");

  CLIENTS_LOCK (sink);
  for (clients = sink->clients; clients; clients = next) {
    GstTCPClient *client;

    client = (GstTCPClient *) clients->data;
    next = g_list_next (clients);

    client->status = GST_CLIENT_STATUS_REMOVED;
    gst_multi_fd_sink_remove_client_link (sink, clients);
  }
  SEND_COMMAND (sink, CONTROL_RESTART);
  CLIENTS_UNLOCK (sink);
}

/* "get-stats" signal implemntation
 * the array returned contains:
 *
 * guint64 : bytes_sent
 * guint64 : connect time (in nanoseconds)
 * guint64 : disconnect time (in nanoseconds)
 * guint64 : time the client is/was connected (in nanoseconds)
 * guint64 : last activity time (in nanoseconds)
 */
GValueArray *
gst_multi_fd_sink_get_stats (GstMultiFdSink * sink, int fd)
{
  GstTCPClient *client;
  GValueArray *result = NULL;
  GList *clink;

  CLIENTS_LOCK (sink);
  clink = g_hash_table_lookup (sink->fd_hash, &fd);
  client = (GstTCPClient *) clink->data;
  if (client != NULL) {
    GValue value = { 0 };
    guint64 interval;

    result = g_value_array_new (5);

    g_value_init (&value, G_TYPE_UINT64);
    g_value_set_uint64 (&value, client->bytes_sent);
    result = g_value_array_append (result, &value);
    g_value_unset (&value);
    g_value_init (&value, G_TYPE_UINT64);
    g_value_set_uint64 (&value, client->connect_time);
    result = g_value_array_append (result, &value);
    g_value_unset (&value);
    if (client->disconnect_time == 0) {
      GTimeVal nowtv;

      g_get_current_time (&nowtv);

      interval = GST_TIMEVAL_TO_TIME (nowtv) - client->connect_time;
    } else {
      interval = client->disconnect_time - client->connect_time;
    }
    g_value_init (&value, G_TYPE_UINT64);
    g_value_set_uint64 (&value, client->disconnect_time);
    result = g_value_array_append (result, &value);
    g_value_unset (&value);
    g_value_init (&value, G_TYPE_UINT64);
    g_value_set_uint64 (&value, interval);
    result = g_value_array_append (result, &value);
    g_value_unset (&value);
    g_value_init (&value, G_TYPE_UINT64);
    g_value_set_uint64 (&value, client->last_activity_time);
    result = g_value_array_append (result, &value);
  }
  CLIENTS_UNLOCK (sink);

  /* python doesn't like a NULL pointer yet */
  if (result == NULL) {
    GST_WARNING_OBJECT (sink, "[fd %5d] no client with this found!", fd);
    result = g_value_array_new (0);
  }

  return result;
}

/* should be called with the clientslock helt.
 * Note that we don't close the fd as we didn't open it in the first
 * place. An application should connect to the client-removed signal and
 * close the fd itself.
 */
static void
gst_multi_fd_sink_remove_client_link (GstMultiFdSink * sink, GList * link)
{
  int fd;
  GTimeVal now;
  GstTCPClient *client = (GstTCPClient *) link->data;
  GstMultiFdSinkClass *fclass;

  fclass = GST_MULTI_FD_SINK_GET_CLASS (sink);

  fd = client->fd.fd;

  /* FIXME: if we keep track of ip we can log it here and signal */
  switch (client->status) {
    case GST_CLIENT_STATUS_OK:
      GST_WARNING_OBJECT (sink, "[fd %5d] removing client %p for no reason",
          fd, client);
      break;
    case GST_CLIENT_STATUS_CLOSED:
      GST_DEBUG_OBJECT (sink, "[fd %5d] removing client %p because of close",
          fd, client);
      break;
    case GST_CLIENT_STATUS_REMOVED:
      GST_DEBUG_OBJECT (sink,
          "[fd %5d] removing client %p because the app removed it", fd, client);
      break;
    case GST_CLIENT_STATUS_SLOW:
      GST_INFO_OBJECT (sink,
          "[fd %5d] removing client %p because it was too slow", fd, client);
      break;
    case GST_CLIENT_STATUS_ERROR:
      GST_WARNING_OBJECT (sink,
          "[fd %5d] removing client %p because of error", fd, client);
      break;
    default:
      GST_WARNING_OBJECT (sink,
          "[fd %5d] removing client %p with invalid reason", fd, client);
      break;
  }

  gst_fdset_remove_fd (sink->fdset, &client->fd);

  g_get_current_time (&now);
  client->disconnect_time = GST_TIMEVAL_TO_TIME (now);

  /* free client buffers */
  g_slist_foreach (client->sending, (GFunc) gst_mini_object_unref, NULL);
  g_slist_free (client->sending);
  client->sending = NULL;

  if (client->caps)
    gst_caps_unref (client->caps);
  client->caps = NULL;

  /* unlock the mutex before signaling because the signal handler
   * might query some properties */
  CLIENTS_UNLOCK (sink);

  g_signal_emit (G_OBJECT (sink),
      gst_multi_fd_sink_signals[SIGNAL_CLIENT_REMOVED], 0, fd, client->status);

  /* lock again before we remove the client completely */
  CLIENTS_LOCK (sink);

  /* fd cannot be reused in the above signal callback so we can safely
   * remove it from the hashtable here */
  if (!g_hash_table_remove (sink->fd_hash, &client->fd.fd)) {
    GST_WARNING_OBJECT (sink,
        "[fd %5d] error removing client %p from hash", client->fd.fd, client);
  }
  /* after releasing the lock above, the link could be invalid, more
   * precisely, the next and prev pointers could point to invalid list
   * links. One optimisation could be to add a cookie to the linked list
   * and take a shortcut when it did not change between unlocking and locking
   * our mutex. For now we just walk the list again. */
  sink->clients = g_list_remove (sink->clients, client);

  if (fclass->removed)
    fclass->removed (sink, client->fd.fd);

  g_free (client);
  CLIENTS_UNLOCK (sink);

  /* and the fd is really gone now */
  g_signal_emit (G_OBJECT (sink),
      gst_multi_fd_sink_signals[SIGNAL_CLIENT_FD_REMOVED], 0, fd);

  CLIENTS_LOCK (sink);
}

/* handle a read on a client fd,
 * which either indicates a close or should be ignored
 * returns FALSE if some error occured or the client closed. */
static gboolean
gst_multi_fd_sink_handle_client_read (GstMultiFdSink * sink,
    GstTCPClient * client)
{
  int avail, fd;
  gboolean ret;

  fd = client->fd.fd;

  if (ioctl (fd, FIONREAD, &avail) < 0)
    goto ioctl_failed;

  GST_DEBUG_OBJECT (sink, "[fd %5d] select reports client read of %d bytes",
      fd, avail);

  ret = TRUE;

  if (avail == 0) {
    /* client sent close, so remove it */
    GST_DEBUG_OBJECT (sink, "[fd %5d] client asked for close, removing", fd);
    client->status = GST_CLIENT_STATUS_CLOSED;
    ret = FALSE;
  } else if (avail < 0) {
    GST_WARNING_OBJECT (sink, "[fd %5d] avail < 0, removing", fd);
    client->status = GST_CLIENT_STATUS_ERROR;
    ret = FALSE;
  } else {
    guint8 dummy[512];
    gint nread;

    /* just Read 'n' Drop, could also just drop the client as it's not supposed
     * to write to us except for closing the socket, I guess it's because we
     * like to listen to our customers. */
    do {
      /* this is the maximum we can read */
      gint to_read = MIN (avail, 512);

      GST_DEBUG_OBJECT (sink, "[fd %5d] client wants us to read %d bytes",
          fd, to_read);

      nread = read (fd, dummy, to_read);
      if (nread < -1) {
        GST_WARNING_OBJECT (sink, "[fd %5d] could not read %d bytes: %s (%d)",
            fd, to_read, g_strerror (errno), errno);
        client->status = GST_CLIENT_STATUS_ERROR;
        ret = FALSE;
        break;
      } else if (nread == 0) {
        GST_WARNING_OBJECT (sink, "[fd %5d] 0 bytes in read, removing", fd);
        client->status = GST_CLIENT_STATUS_ERROR;
        ret = FALSE;
        break;
      }
      avail -= nread;
    }
    while (avail > 0);
  }
  return ret;

  /* ERRORS */
ioctl_failed:
  {
    GST_WARNING_OBJECT (sink, "[fd %5d] ioctl failed: %s (%d)",
        fd, g_strerror (errno), errno);
    client->status = GST_CLIENT_STATUS_ERROR;
    return FALSE;
  }
}

/* Queue raw data for this client, creating a new buffer.
 * This takes ownership of the data by
 * setting it as GST_BUFFER_MALLOCDATA() on the created buffer so
 * be sure to pass g_free()-able @data.
 */
static gboolean
gst_multi_fd_sink_client_queue_data (GstMultiFdSink * sink,
    GstTCPClient * client, gchar * data, gint len)
{
  GstBuffer *buf;

  buf = gst_buffer_new ();
  GST_BUFFER_DATA (buf) = (guint8 *) data;
  GST_BUFFER_MALLOCDATA (buf) = (guint8 *) data;
  GST_BUFFER_SIZE (buf) = len;

  GST_LOG_OBJECT (sink, "[fd %5d] queueing data of length %d",
      client->fd.fd, len);

  client->sending = g_slist_append (client->sending, buf);

  return TRUE;
}

/* GDP-encode given caps and queue them for sending */
static gboolean
gst_multi_fd_sink_client_queue_caps (GstMultiFdSink * sink,
    GstTCPClient * client, const GstCaps * caps)
{
  guint8 *header;
  guint8 *payload;
  guint length;
  gchar *string;

  g_return_val_if_fail (caps != NULL, FALSE);

  string = gst_caps_to_string (caps);
  GST_DEBUG_OBJECT (sink, "[fd %5d] Queueing caps %s through GDP",
      client->fd.fd, string);
  g_free (string);

  if (!gst_dp_packet_from_caps (caps, sink->header_flags, &length, &header,
          &payload)) {
    GST_DEBUG_OBJECT (sink, "Could not create GDP packet from caps");
    return FALSE;
  }
  gst_multi_fd_sink_client_queue_data (sink, client, (gchar *) header, length);

  length = gst_dp_header_payload_length (header);
  gst_multi_fd_sink_client_queue_data (sink, client, (gchar *) payload, length);

  return TRUE;
}

static gboolean
is_sync_frame (GstMultiFdSink * sink, GstBuffer * buffer)
{
  if (GST_BUFFER_FLAG_IS_SET (buffer, GST_BUFFER_FLAG_DELTA_UNIT)) {
    return FALSE;
  } else if (!GST_BUFFER_FLAG_IS_SET (buffer, GST_BUFFER_FLAG_IN_CAPS)) {
    return TRUE;
  }
  return FALSE;
}

/* queue the given buffer for the given client, possibly adding the GDP
 * header if GDP is being used */
static gboolean
gst_multi_fd_sink_client_queue_buffer (GstMultiFdSink * sink,
    GstTCPClient * client, GstBuffer * buffer)
{
  GstCaps *caps;

  /* TRUE: send them if the new caps have them */
  gboolean send_streamheader = FALSE;
  GstStructure *s;


  /* before we queue the buffer, we check if we need to queue streamheader
   * buffers (because it's a new client, or because they changed) */
  caps = gst_buffer_get_caps (buffer);  /* cleaned up after streamheader */
  if (!client->caps) {
    GST_LOG_OBJECT (sink,
        "[fd %5d] no previous caps for this client, send streamheader",
        client->fd.fd);
    send_streamheader = TRUE;
    client->caps = gst_caps_ref (caps);
  } else {
    /* there were previous caps recorded, so compare */
    if (!gst_caps_is_equal (caps, client->caps)) {
      const GValue *sh1, *sh2;

      /* caps are not equal, but could still have the same streamheader */
      s = gst_caps_get_structure (caps, 0);
      if (!gst_structure_has_field (s, "streamheader")) {
        /* no new streamheader, so nothing new to send */
        GST_LOG_OBJECT (sink,
            "[fd %5d] new caps do not have streamheader, not sending",
            client->fd.fd);
      } else {
        /* there is a new streamheader */
        s = gst_caps_get_structure (client->caps, 0);
        if (!gst_structure_has_field (s, "streamheader")) {
          /* no previous streamheader, so send the new one */
          GST_LOG_OBJECT (sink,
              "[fd %5d] previous caps did not have streamheader, sending",
              client->fd.fd);
          send_streamheader = TRUE;
        } else {
          /* both old and new caps have streamheader set */
          sh1 = gst_structure_get_value (s, "streamheader");
          s = gst_caps_get_structure (caps, 0);
          sh2 = gst_structure_get_value (s, "streamheader");
          if (gst_value_compare (sh1, sh2) != GST_VALUE_EQUAL) {
            GST_LOG_OBJECT (sink,
                "[fd %5d] new streamheader different from old, sending",
                client->fd.fd);
            send_streamheader = TRUE;
          }
        }
      }
    }
  }

  if (G_UNLIKELY (send_streamheader)) {
    const GValue *sh;
    GArray *buffers;
    int i;

    GST_LOG_OBJECT (sink,
        "[fd %5d] sending streamheader from caps %" GST_PTR_FORMAT,
        client->fd.fd, caps);
    s = gst_caps_get_structure (caps, 0);
    if (!gst_structure_has_field (s, "streamheader")) {
      GST_LOG_OBJECT (sink,
          "[fd %5d] no new streamheader, so nothing to send", client->fd.fd);
    } else {
      GST_LOG_OBJECT (sink,
          "[fd %5d] sending streamheader from caps %" GST_PTR_FORMAT,
          client->fd.fd, caps);
      sh = gst_structure_get_value (s, "streamheader");
      g_assert (G_VALUE_TYPE (sh) == GST_TYPE_ARRAY);
      buffers = g_value_peek_pointer (sh);
      for (i = 0; i < buffers->len; ++i) {
        GValue *bufval;
        GstBuffer *buffer;

        bufval = &g_array_index (buffers, GValue, i);
        g_assert (G_VALUE_TYPE (bufval) == GST_TYPE_BUFFER);
        buffer = g_value_peek_pointer (bufval);
        GST_LOG_OBJECT (sink,
            "[fd %5d] queueing streamheader buffer of length %d",
            client->fd.fd, GST_BUFFER_SIZE (buffer));
        gst_buffer_ref (buffer);

        if (sink->protocol == GST_TCP_PROTOCOL_GDP) {
          guint8 *header;
          guint len;

          if (!gst_dp_header_from_buffer (buffer, sink->header_flags, &len,
                  &header)) {
            GST_DEBUG_OBJECT (sink,
                "[fd %5d] could not create header, removing client",
                client->fd.fd);
            return FALSE;
          }
          gst_multi_fd_sink_client_queue_data (sink, client, (gchar *) header,
              len);
        }

        client->sending = g_slist_append (client->sending, buffer);
      }
    }
  }

  gst_caps_unref (caps);
  caps = NULL;
  /* now we can send the buffer, possibly sending a GDP header first */
  if (sink->protocol == GST_TCP_PROTOCOL_GDP) {
    guint8 *header;
    guint len;

    if (!gst_dp_header_from_buffer (buffer, sink->header_flags, &len, &header)) {
      GST_DEBUG_OBJECT (sink,
          "[fd %5d] could not create header, removing client", client->fd.fd);
      return FALSE;
    }
    gst_multi_fd_sink_client_queue_data (sink, client, (gchar *) header, len);
  }

  GST_LOG_OBJECT (sink, "[fd %5d] queueing buffer of length %d",
      client->fd.fd, GST_BUFFER_SIZE (buffer));

  gst_buffer_ref (buffer);
  client->sending = g_slist_append (client->sending, buffer);

  return TRUE;
}

/* find the keyframe in the list of buffers starting the
 * search from @idx. @direction as -1 will search backwards, 
 * 1 will search forwards.
 * Returns: the index or -1 if there is no keyframe after idx.
 */
static gint
find_syncframe (GstMultiFdSink * sink, gint idx, gint direction)
{
  gint i, len, result;

  /* take length of queued buffers */
  len = sink->bufqueue->len;

  /* assume we don't find a keyframe */
  result = -1;

  /* then loop over all buffers to find the first keyframe */
  for (i = idx; i >= 0 && i < len; i += direction) {
    GstBuffer *buf;

    buf = g_array_index (sink->bufqueue, GstBuffer *, i);
    if (is_sync_frame (sink, buf)) {
      GST_LOG_OBJECT (sink, "found keyframe at %d from %d, direction %d",
          i, idx, direction);
      result = i;
      break;
    }
  }
  return result;
}

#define find_next_syncframe(s,i) 	find_syncframe(s,i,1)
#define find_prev_syncframe(s,i) 	find_syncframe(s,i,-1)

/* find the positions in the buffer queue where *_min and *_max
 * is satisfied
 */
/* count the amount of data in the buffers and return the index
 * that satifies the given limits.
 *
 * Returns: index @idx in the buffer queue so that the given limits are
 * satisfied. TRUE if all the limits could be satisfied, FALSE if not
 * enough data was in the queue.
 *
 * FIXME, this code might now work if any of the units is in buffers...
 */
static gboolean
find_limits (GstMultiFdSink * sink,
    gint * min_idx, gint bytes_min, gint buffers_min, gint64 time_min,
    gint * max_idx, gint bytes_max, gint buffers_max, gint64 time_max)
{
  GstClockTime first, time;
  gint i, len, bytes;
  gboolean result, max_hit;

  /* take length of queue */
  len = sink->bufqueue->len;

  /* this must hold */
  g_assert (len > 0);

  GST_LOG_OBJECT (sink,
      "bytes_min %d, buffers_min %d, time_min %" GST_TIME_FORMAT
      ", bytes_max %d, buffers_max %d, time_max %" GST_TIME_FORMAT, bytes_min,
      buffers_min, GST_TIME_ARGS (time_min), bytes_max, buffers_max,
      GST_TIME_ARGS (time_max));

  /* do the trivial buffer limit test */
  if (buffers_min != -1 && len < buffers_min) {
    *min_idx = len - 1;
    *max_idx = len - 1;
    return FALSE;
  }

  result = FALSE;
  /* else count bytes and time */
  first = -1;
  bytes = 0;
  /* unset limits */
  *min_idx = -1;
  *max_idx = -1;
  max_hit = FALSE;

  i = 0;
  /* loop through the buffers, when a limit is ok, mark it 
   * as -1, we have at least one buffer in the queue. */
  do {
    GstBuffer *buf;

    /* if we checked all min limits, update result */
    if (bytes_min == -1 && time_min == -1 && *min_idx == -1) {
      /* don't go below 0 */
      *min_idx = MAX (i - 1, 0);
    }
    /* if we reached one max limit break out */
    if (max_hit) {
      /* i > 0 when we get here, we subtract one to get the position
       * of the previous buffer. */
      *max_idx = i - 1;
      /* we have valid complete result if we found a min_idx too */
      result = *min_idx != -1;
      break;
    }
    buf = g_array_index (sink->bufqueue, GstBuffer *, i);

    bytes += GST_BUFFER_SIZE (buf);

    /* take timestamp and save for the base first timestamp */
    if ((time = GST_BUFFER_TIMESTAMP (buf)) != -1) {
      if (first == -1)
        first = time;

      /* increase max usage if we did not fill enough. Note that
       * buffers are sorted from new to old, so the first timestamp is
       * bigger than the next one. */
      if (time_min != -1 && first - time >= time_min)
        time_min = -1;
      if (time_max != -1 && first - time >= time_max)
        max_hit = TRUE;
    }
    /* time is OK or unknown, check and increase if not enough bytes */
    if (bytes_min != -1) {
      if (bytes >= bytes_min)
        bytes_min = -1;
    }
    if (bytes_max != -1) {
      if (bytes >= bytes_max) {
        max_hit = TRUE;
      }
    }
    i++;
  }
  while (i < len);

  /* if we did not hit the max or min limit, set to buffer size */
  if (*max_idx == -1)
    *max_idx = len - 1;
  /* make sure min does not exceed max */
  if (*min_idx == -1)
    *min_idx = *max_idx;

  return result;
}

/* parse the unit/value pair and assign it to the result value of the
 * right type, leave the other values untouched 
 *
 * Returns: FALSE if the unit is unknown or undefined. TRUE otherwise.
 */
static gboolean
assign_value (GstUnitType unit, guint64 value, gint * bytes, gint * buffers,
    GstClockTime * time)
{
  gboolean res = TRUE;

  /* set only the limit of the given format to the given value */
  switch (unit) {
    case GST_UNIT_TYPE_BUFFERS:
      *buffers = (gint) value;
      break;
    case GST_UNIT_TYPE_TIME:
      *time = value;
      break;
    case GST_UNIT_TYPE_BYTES:
      *bytes = (gint) value;
      break;
    case GST_UNIT_TYPE_UNDEFINED:
    default:
      res = FALSE;
      break;
  }
  return res;
}

/* count the index in the buffer queue to satisfy the given unit
 * and value pair starting from buffer at index 0.
 *
 * Returns: TRUE if there was enuough data in the queue to satisfy the
 * burst values. @idx contains the index in the buffer that contains enough
 * data to satisfy the limits or the last buffer in the queue when the
 * function returns FALSE.
 */
static gboolean
count_burst_unit (GstMultiFdSink * sink, gint * min_idx, GstUnitType min_unit,
    guint64 min_value, gint * max_idx, GstUnitType max_unit, guint64 max_value)
{
  gint bytes_min = -1, buffers_min = -1;
  gint bytes_max = -1, buffers_max = -1;
  GstClockTime time_min = -1, time_max = -1;

  assign_value (min_unit, min_value, &bytes_min, &buffers_min, &time_min);
  assign_value (max_unit, max_value, &bytes_max, &buffers_max, &time_max);

  return find_limits (sink, min_idx, bytes_min, buffers_min, time_min,
      max_idx, bytes_max, buffers_max, time_max);
}

/* decide where in the current buffer queue this new client should start
 * receiving buffers from.
 * This function is called whenever a client is connected and has not yet
 * received a buffer.
 * If this returns -1, it means that we haven't found a good point to
 * start streaming from yet, and this function should be called again later
 * when more buffers have arrived.
 */
static gint
gst_multi_fd_sink_new_client (GstMultiFdSink * sink, GstTCPClient * client)
{
  gint result;

  switch (client->sync_method) {
    case GST_SYNC_METHOD_LATEST:
      /* no syncing, we are happy with whatever the client is going to get */
      GST_LOG_OBJECT (sink, "no client sync needed");
      result = client->bufpos;
      break;
    case GST_SYNC_METHOD_NEXT_KEYFRAME:
    {
      GstBuffer *buf;

      /* if the buffer at the head of the queue is a sync point we can proceed,
       * else we need to skip the buffer and wait for a new one */
      GST_LOG_OBJECT (sink,
          "[fd %5d] new client, bufpos %d, waiting for keyframe", client->fd.fd,
          client->bufpos);

      /* get the buffer for the client */
      buf = g_array_index (sink->bufqueue, GstBuffer *, 0);
      if (is_sync_frame (sink, buf)) {
        GST_LOG_OBJECT (sink, "[fd %5d] new client, found sync", client->fd.fd);
        result = 0;
        goto done;
      }
      /* client is not on a syncbuffer, need to skip this buffer and
       * wait some more */
      GST_LOG_OBJECT (sink, "[fd %5d] new client, skipping buffer",
          client->fd.fd);
      client->bufpos = -1;
      result = -1;
      break;
    }
    case GST_SYNC_METHOD_LATEST_KEYFRAME:
    {
      GST_LOG_OBJECT (sink, "[fd %5d] new client, bufpos %d, bursting keyframe",
          client->fd.fd, client->bufpos);

      /* for new clients we initially scan the complete buffer queue for 
       * a sync point when a buffer is added. If we don't find a keyframe,
       * we need to wait for the next keyframe and so we change the client's
       * sync method to GST_SYNC_METHOD_NEXT_KEYFRAME.
       */
      result = find_next_syncframe (sink, 0);
      if (result != -1)
        goto done;

      GST_LOG_OBJECT (sink, "no keyframe found");
      /* throw client to the waiting state */
      client->bufpos = -1;
      /* and make client sync to next keyframe */
      client->sync_method = GST_SYNC_METHOD_NEXT_KEYFRAME;
      break;
    }
    case GST_SYNC_METHOD_BURST:
    {
      gboolean ok;
      gint max;

      /* move to the position where we satisfy the client's burst
       * parameters. If we could not satisfy the parameters because there
       * is not enough data, we just send what we have (which is in result).
       * We use the max value to limit the search 
       */
      ok = count_burst_unit (sink, &result, client->burst_min_unit,
          client->burst_min_value, &max, client->burst_max_unit,
          client->burst_max_value);

      GST_LOG_OBJECT (sink, "min %d, max %d", result, max);

      /* we hit the max and it is below the min, use that then */
      if (max != -1 && max <= result) {
        result = MAX (max - 1, 0);
      }
      break;
    }
    case GST_SYNC_METHOD_BURST_KEYFRAME:
    {
      gboolean ok;
      gint min_idx, max_idx;
      gint next_syncframe, prev_syncframe;

      /* BURST_KEYFRAME:
       *
       * _always_ start sending a keyframe to the client. We first search
       * a keyframe between min/max limits. If there is none, we send it the
       * last keyframe before min. If there is none, the behaviour is like
       * NEXT_KEYFRAME.
       */
      /* gather burst limits */
      ok = count_burst_unit (sink, &min_idx, client->burst_min_unit,
          client->burst_min_value, &max_idx, client->burst_max_unit,
          client->burst_max_value);

      GST_LOG_OBJECT (sink, "min %d, max %d", min_idx, max_idx);

      /* first find a keyframe after min_idx */
      next_syncframe = find_next_syncframe (sink, min_idx);
      if (next_syncframe != -1 && next_syncframe < max_idx) {
        /* we have a valid keyframe and it's below the max */
        GST_LOG_OBJECT (sink, "found keyframe in min/max limits");
        result = next_syncframe;
        break;
      }

      /* no valid keyframe, try to find one below min */
      prev_syncframe = find_prev_syncframe (sink, min_idx);
      if (prev_syncframe != -1) {
        GST_WARNING_OBJECT (sink,
            "using keyframe below min in BURST_KEYFRAME sync mode");
        result = prev_syncframe;
        break;
      }

      /* no prev keyframe or not enough data  */
      GST_WARNING_OBJECT (sink,
          "no prev keyframe found in BURST_KEYFRAME sync mode, waiting for next");

      /* throw client to the waiting state */
      client->bufpos = -1;
      /* and make client sync to next keyframe */
      client->sync_method = GST_SYNC_METHOD_NEXT_KEYFRAME;
      result = -1;
      break;
    }
    case GST_SYNC_METHOD_BURST_WITH_KEYFRAME:
    {
      gboolean ok;
      gint min_idx, max_idx;
      gint next_syncframe;

      /* BURST_WITH_KEYFRAME:
       *
       * try to start sending a keyframe to the client. We first search
       * a keyframe between min/max limits. If there is none, we send it the
       * amount of data up 'till min.
       */
      /* gather enough data to burst */
      ok = count_burst_unit (sink, &min_idx, client->burst_min_unit,
          client->burst_min_value, &max_idx, client->burst_max_unit,
          client->burst_max_value);

      GST_LOG_OBJECT (sink, "min %d, max %d", min_idx, max_idx);

      /* first find a keyframe after min_idx */
      next_syncframe = find_next_syncframe (sink, min_idx);
      if (next_syncframe != -1 && next_syncframe < max_idx) {
        /* we have a valid keyframe and it's below the max */
        GST_LOG_OBJECT (sink, "found keyframe in min/max limits");
        result = next_syncframe;
        break;
      }

      /* no keyframe, send data from min_idx */
      GST_WARNING_OBJECT (sink, "using min in BURST_WITH_KEYFRAME sync mode");

      /* make sure we don't go over the max limit */
      if (max_idx != -1 && max_idx <= min_idx) {
        result = MAX (max_idx - 1, 0);
      } else {
        result = min_idx;
      }

      break;
    }
    default:
      g_warning ("unknown sync method %d", client->sync_method);
      result = client->bufpos;
      break;
  }
done:
  return result;
}

/* Handle a write on a client,
 * which indicates a read request from a client.
 *
 * For each client we maintain a queue of GstBuffers that contain the raw bytes
 * we need to send to the client. In the case of the GDP protocol, we create
 * buffers out of the header bytes so that we can focus only on sending
 * buffers.
 *
 * We first check to see if we need to send caps (in GDP) and streamheaders.
 * If so, we queue them.
 *
 * Then we run into the main loop that tries to send as many buffers as
 * possible. It will first exhaust the client->sending queue and if the queue
 * is empty, it will pick a buffer from the global queue.
 *
 * Sending the buffers from the client->sending queue is basically writing
 * the bytes to the socket and maintaining a count of the bytes that were
 * sent. When the buffer is completely sent, it is removed from the
 * client->sending queue and we try to pick a new buffer for sending.
 *
 * When the sending returns a partial buffer we stop sending more data as
 * the next send operation could block.
 *
 * This functions returns FALSE if some error occured.
 */
static gboolean
gst_multi_fd_sink_handle_client_write (GstMultiFdSink * sink,
    GstTCPClient * client)
{
  int fd = client->fd.fd;
  gboolean more;
  gboolean res;
  GstClockTime now;
  GTimeVal nowtv;

  g_get_current_time (&nowtv);
  now = GST_TIMEVAL_TO_TIME (nowtv);

  /* when using GDP, first check if we have queued caps yet */
  if (sink->protocol == GST_TCP_PROTOCOL_GDP) {
    if (!client->caps_sent) {
      GstPad *peer;
      GstCaps *caps;

      peer = gst_pad_get_peer (GST_BASE_SINK_PAD (sink));
      if (!peer) {
        GST_WARNING_OBJECT (sink, "pad has no peer");
        return FALSE;
      }
      gst_object_unref (peer);

      caps = gst_pad_get_negotiated_caps (GST_BASE_SINK_PAD (sink));
      if (!caps) {
        GST_WARNING_OBJECT (sink, "pad caps not yet negotiated");
        return FALSE;
      }

      /* queue caps for sending */
      res = gst_multi_fd_sink_client_queue_caps (sink, client, caps);

      gst_caps_unref (caps);

      if (!res) {
        GST_DEBUG_OBJECT (sink, "Failed queueing caps, removing client");
        return FALSE;
      }
      client->caps_sent = TRUE;
    }
  }

  more = TRUE;
  do {
    gint maxsize;

    if (!client->sending) {
      /* client is not working on a buffer */
      if (client->bufpos == -1) {
        /* client is too fast, remove from write queue until new buffer is
         * available */
        gst_fdset_fd_ctl_write (sink->fdset, &client->fd, FALSE);
        return TRUE;
      } else {
        /* client can pick a buffer from the global queue */
        GstBuffer *buf;

        /* for new connections, we need to find a good spot in the
         * bufqueue to start streaming from */
        if (client->new_connection) {
          gint position = gst_multi_fd_sink_new_client (sink, client);

          if (position >= 0) {
            /* we got a valid spot in the queue */
            client->new_connection = FALSE;
            client->bufpos = position;
          } else {
            /* cannot send data to this client yet */
            gst_fdset_fd_ctl_write (sink->fdset, &client->fd, FALSE);
            return TRUE;
          }
        }

        /* grab buffer */
        buf = g_array_index (sink->bufqueue, GstBuffer *, client->bufpos);
        client->bufpos--;
        GST_LOG_OBJECT (sink, "[fd %5d] client %p at position %d",
            fd, client, client->bufpos);

        /* queueing a buffer will ref it */
        gst_multi_fd_sink_client_queue_buffer (sink, client, buf);

        /* need to start from the first byte for this new buffer */
        client->bufoffset = 0;
      }
    }

    /* see if we need to send something */
    if (client->sending) {
      ssize_t wrote;
      GstBuffer *head;

      /* pick first buffer from list */
      head = GST_BUFFER (client->sending->data);
      maxsize = GST_BUFFER_SIZE (head) - client->bufoffset;

      /* try to write the complete buffer */
#ifdef MSG_NOSIGNAL
#define FLAGS MSG_NOSIGNAL
#else
#define FLAGS 0
#endif
      if (client->is_socket) {
        wrote =
            send (fd, GST_BUFFER_DATA (head) + client->bufoffset, maxsize,
            FLAGS);
      } else {
        wrote = write (fd, GST_BUFFER_DATA (head) + client->bufoffset, maxsize);
      }

      if (wrote < 0) {
        /* hmm error.. */
        if (errno == EAGAIN) {
          /* nothing serious, resource was unavailable, try again later */
          more = FALSE;
        } else if (errno == ECONNRESET) {
          goto connection_reset;
        } else {
          goto write_error;
        }
      } else {
        if (wrote < maxsize) {
          /* partial write means that the client cannot read more and we should
           * stop sending more */
          GST_LOG_OBJECT (sink, "partial write on %d of %d bytes", fd, wrote);
          client->bufoffset += wrote;
          more = FALSE;
        } else {
          /* complete buffer was written, we can proceed to the next one */
          client->sending = g_slist_remove (client->sending, head);
          gst_buffer_unref (head);
          /* make sure we start from byte 0 for the next buffer */
          client->bufoffset = 0;
        }
        /* update stats */
        client->bytes_sent += wrote;
        client->last_activity_time = now;
        sink->bytes_served += wrote;
      }
    }
  } while (more);

  return TRUE;

  /* ERRORS */
connection_reset:
  {
    GST_DEBUG_OBJECT (sink, "[fd %5d] connection reset by peer, removing", fd);
    client->status = GST_CLIENT_STATUS_CLOSED;
    return FALSE;
  }
write_error:
  {
    GST_WARNING_OBJECT (sink,
        "[fd %5d] could not write, removing client: %s (%d)", fd,
        g_strerror (errno), errno);
    client->status = GST_CLIENT_STATUS_ERROR;
    return FALSE;
  }
}

/* calculate the new position for a client after recovery. This function
 * does not update the client position but merely returns the required
 * position.
 */
static gint
gst_multi_fd_sink_recover_client (GstMultiFdSink * sink, GstTCPClient * client)
{
  gint newbufpos;

  GST_WARNING_OBJECT (sink,
      "[fd %5d] client %p is lagging at %d, recover using policy %d",
      client->fd.fd, client, client->bufpos, sink->recover_policy);

  switch (sink->recover_policy) {
    case GST_RECOVER_POLICY_NONE:
      /* do nothing, client will catch up or get kicked out when it reaches
       * the hard max */
      newbufpos = client->bufpos;
      break;
    case GST_RECOVER_POLICY_RESYNC_LATEST:
      /* move to beginning of queue */
      newbufpos = -1;
      break;
    case GST_RECOVER_POLICY_RESYNC_SOFT_LIMIT:
      /* move to beginning of soft max */
      newbufpos = sink->units_soft_max;
      break;
    case GST_RECOVER_POLICY_RESYNC_KEYFRAME:
      /* find keyframe in buffers, we search backwards to find the
       * closest keyframe relative to what this client already received. */
      newbufpos = MIN (sink->bufqueue->len - 1, sink->units_soft_max - 1);

      while (newbufpos >= 0) {
        GstBuffer *buf;

        buf = g_array_index (sink->bufqueue, GstBuffer *, newbufpos);
        if (is_sync_frame (sink, buf)) {
          /* found a buffer that is not a delta unit */
          break;
        }
        newbufpos--;
      }
      break;
    default:
      /* unknown recovery procedure */
      newbufpos = sink->units_soft_max;
      break;
  }
  return newbufpos;
}

/* Queue a buffer on the global queue.
 *
 * This function adds the buffer to the front of a GArray. It removes the
 * tail buffer if the max queue size is exceeded, unreffing the queued buffer.
 * Note that unreffing the buffer is not a problem as clients who
 * started writing out this buffer will still have a reference to it in the
 * client->sending queue.
 *
 * After adding the buffer, we update all client positions in the queue. If
 * a client moves over the soft max, we start the recovery procedure for this
 * slow client. If it goes over the hard max, it is put into the slow list
 * and removed.
 *
 * Special care is taken of clients that were waiting for a new buffer (they
 * had a position of -1) because they can proceed after adding this new buffer.
 * This is done by adding the client back into the write fd_set and signalling
 * the select thread that the fd_set changed.
 */
static void
gst_multi_fd_sink_queue_buffer (GstMultiFdSink * sink, GstBuffer * buf)
{
  GList *clients, *next;
  gint queuelen;
  gboolean need_signal = FALSE;
  gint max_buffer_usage;
  gint i;
  GTimeVal nowtv;
  GstClockTime now;

  g_get_current_time (&nowtv);
  now = GST_TIMEVAL_TO_TIME (nowtv);

  CLIENTS_LOCK (sink);
  /* add buffer to queue */
  g_array_prepend_val (sink->bufqueue, buf);
  queuelen = sink->bufqueue->len;

  /* then loop over the clients and update the positions */
  max_buffer_usage = 0;
  for (clients = sink->clients; clients; clients = next) {
    GstTCPClient *client;

    client = (GstTCPClient *) clients->data;
    next = g_list_next (clients);

    client->bufpos++;
    GST_LOG_OBJECT (sink, "[fd %5d] client %p at position %d",
        client->fd.fd, client, client->bufpos);
    /* check soft max if needed, recover client */
    if (sink->units_soft_max > 0 && client->bufpos >= sink->units_soft_max) {
      gint newpos;

      newpos = gst_multi_fd_sink_recover_client (sink, client);
      if (newpos != client->bufpos) {
        client->bufpos = newpos;
        client->discont = TRUE;
        GST_INFO_OBJECT (sink, "[fd %5d] client %p position reset to %d",
            client->fd.fd, client, client->bufpos);
      } else {
        GST_INFO_OBJECT (sink,
            "[fd %5d] client %p not recovering position",
            client->fd.fd, client);
      }
    }
    /* check hard max and timeout, remove client */
    if ((sink->units_max > 0 && client->bufpos >= sink->units_max) ||
        (sink->timeout > 0
            && now - client->last_activity_time > sink->timeout)) {
      /* remove client */
      GST_WARNING_OBJECT (sink, "[fd %5d] client %p is too slow, removing",
          client->fd.fd, client);
      /* remove the client, the fd set will be cleared and the select thread
       * will be signaled */
      client->status = GST_CLIENT_STATUS_SLOW;
      gst_multi_fd_sink_remove_client_link (sink, clients);
      /* set client to invalid position while being removed */
      client->bufpos = -1;
      need_signal = TRUE;
    } else if (client->bufpos == 0 || client->new_connection) {
      /* can send data to this client now. need to signal the select thread that
       * the fd_set changed */
      gst_fdset_fd_ctl_write (sink->fdset, &client->fd, TRUE);
      need_signal = TRUE;
    }
    /* keep track of maximum buffer usage */
    if (client->bufpos > max_buffer_usage) {
      max_buffer_usage = client->bufpos;
    }
  }

  /* make sure we respect bytes-min, buffers-min and time-min when they are set */
  {
    gint usage, max;

    GST_LOG_OBJECT (sink,
        "extending queue %d to respect time_min %" GST_TIME_FORMAT
        ", bytes_min %d, buffers_min %d", max_buffer_usage,
        GST_TIME_ARGS (sink->time_min), sink->bytes_min, sink->buffers_min);

    /* get index where the limits are ok, we don't really care if all limits
     * are ok, we just queue as much as we need. We also don't compare against
     * the max limits. */
    find_limits (sink, &usage, sink->bytes_min, sink->buffers_min,
        sink->time_min, &max, -1, -1, -1);

    max_buffer_usage = MAX (max_buffer_usage, usage + 1);
    GST_LOG_OBJECT (sink, "extended queue to %d", max_buffer_usage);
  }

  /* now look for sync points and make sure there is at least one
   * sync point in the queue. We only do this if the LATEST_KEYFRAME
   * mode is selected */
  if (sink->def_sync_method == GST_SYNC_METHOD_LATEST_KEYFRAME) {
    /* no point in searching beyond the queue length */
    gint limit = queuelen;
    GstBuffer *buf;

    /* no point in searching beyond the soft-max if any. */
    if (sink->units_soft_max > 0) {
      limit = MIN (limit, sink->units_soft_max);
    }
    GST_LOG_OBJECT (sink, "extending queue to include sync point, now at %d",
        max_buffer_usage);
    for (i = 0; i < limit; i++) {
      buf = g_array_index (sink->bufqueue, GstBuffer *, i);
      if (is_sync_frame (sink, buf)) {
        /* found a sync frame, now extend the buffer usage to
         * include at least this frame. */
        max_buffer_usage = MAX (max_buffer_usage, i);
        break;
      }
    }
    GST_LOG_OBJECT (sink, "max buffer usage is now %d", max_buffer_usage);
  }

  GST_LOG_OBJECT (sink, "len %d, usage %d", queuelen, max_buffer_usage);

  /* nobody is referencing units after max_buffer_usage so we can
   * remove them from the queue. We remove them in reverse order as
   * this is the most optimal for GArray. */
  for (i = queuelen - 1; i > max_buffer_usage; i--) {
    GstBuffer *old;

    /* queue exceeded max size */
    queuelen--;
    old = g_array_index (sink->bufqueue, GstBuffer *, i);
    sink->bufqueue = g_array_remove_index (sink->bufqueue, i);

    /* unref tail buffer */
    gst_buffer_unref (old);
  }
  /* save for stats */
  sink->buffers_queued = max_buffer_usage;
  CLIENTS_UNLOCK (sink);

  /* and send a signal to thread if fd_set changed */
  if (need_signal) {
    SEND_COMMAND (sink, CONTROL_RESTART);
  }
}

/* Handle the clients. Basically does a blocking select for one
 * of the client fds to become read or writable. We also have a
 * filedescriptor to receive commands on that we need to check.
 *
 * After going out of the select call, we read and write to all
 * clients that can do so. Badly behaving clients are put on a
 * garbage list and removed.
 */
static void
gst_multi_fd_sink_handle_clients (GstMultiFdSink * sink)
{
  int result;
  GList *clients, *next;
  gboolean try_again;
  GstMultiFdSinkClass *fclass;

  fclass = GST_MULTI_FD_SINK_GET_CLASS (sink);

  do {
    gboolean stop = FALSE;

    try_again = FALSE;

    /* check for:
     * - server socket input (ie, new client connections)
     * - client socket input (ie, clients saying goodbye)
     * - client socket output (ie, client reads)          */
    GST_LOG_OBJECT (sink, "waiting on action on fdset");
    result = gst_fdset_wait (sink->fdset, -1);

    /* < 0 is an error, 0 just means a timeout happened, which is impossible */
    if (result < 0) {
      GST_WARNING_OBJECT (sink, "wait failed: %s (%d)", g_strerror (errno),
          errno);
      if (errno == EBADF) {
        /* ok, so one or more of the fds is invalid. We loop over them to find
         * the ones that give an error to the F_GETFL fcntl. */
        CLIENTS_LOCK (sink);
        for (clients = sink->clients; clients; clients = next) {
          GstTCPClient *client;
          int fd;
          long flags;
          int res;

          client = (GstTCPClient *) clients->data;
          next = g_list_next (clients);

          fd = client->fd.fd;

          res = fcntl (fd, F_GETFL, &flags);
          if (res == -1) {
            GST_WARNING_OBJECT (sink, "fnctl failed for %d, removing: %s (%d)",
                fd, g_strerror (errno), errno);
            if (errno == EBADF) {
              client->status = GST_CLIENT_STATUS_ERROR;
              gst_multi_fd_sink_remove_client_link (sink, clients);
            }
          }
        }
        CLIENTS_UNLOCK (sink);
        /* after this, go back in the select loop as the read/writefds
         * are not valid */
        try_again = TRUE;
      } else if (errno == EINTR) {
        /* interrupted system call, just redo the select */
        try_again = TRUE;
      } else {
        /* this is quite bad... */
        GST_ELEMENT_ERROR (sink, RESOURCE, READ, (NULL),
            ("select failed: %s (%d)", g_strerror (errno), errno));
        return;
      }
    } else {
      GST_LOG_OBJECT (sink, "wait done: %d sockets with events", result);
      /* read all commands */
      if (gst_fdset_fd_can_read (sink->fdset, &READ_SOCKET (sink))) {
        GST_LOG_OBJECT (sink, "have a command");
        while (TRUE) {
          gchar command;
          int res;

          READ_COMMAND (sink, command, res);
          if (res < 0) {
            GST_LOG_OBJECT (sink, "no more commands");
            /* no more commands */
            break;
          }

          switch (command) {
            case CONTROL_RESTART:
              GST_LOG_OBJECT (sink, "restart");
              /* need to restart the select call as the fd_set changed */
              /* if other file descriptors than the READ_SOCKET had activity,
               * we don't restart just yet, but handle the other clients first
               */
              if (result == 1)
                try_again = TRUE;
              break;
            case CONTROL_STOP:
              /* break out of the select loop */
              GST_LOG_OBJECT (sink, "stop");
              /* stop this function */
              stop = TRUE;
              break;
            default:
              GST_WARNING_OBJECT (sink, "unkown");
              g_warning ("multifdsink: unknown control message received");
              break;
          }
        }
      }
    }
    if (stop) {
      return;
    }
  } while (try_again);

  /* subclasses can check fdset with this virtual function */
  if (fclass->wait)
    fclass->wait (sink, sink->fdset);

  /* Check the clients */
  CLIENTS_LOCK (sink);
  for (clients = sink->clients; clients; clients = next) {
    GstTCPClient *client;

    client = (GstTCPClient *) clients->data;
    next = g_list_next (clients);

    if (client->status != GST_CLIENT_STATUS_OK) {
      gst_multi_fd_sink_remove_client_link (sink, clients);
      continue;
    }

    if (gst_fdset_fd_has_closed (sink->fdset, &client->fd)) {
      client->status = GST_CLIENT_STATUS_CLOSED;
      gst_multi_fd_sink_remove_client_link (sink, clients);
      continue;
    }
    if (gst_fdset_fd_has_error (sink->fdset, &client->fd)) {
      GST_WARNING_OBJECT (sink, "gst_fdset_fd_has_error for %d", client->fd);
      client->status = GST_CLIENT_STATUS_ERROR;
      gst_multi_fd_sink_remove_client_link (sink, clients);
      continue;
    }
    if (gst_fdset_fd_can_read (sink->fdset, &client->fd)) {
      /* handle client read */
      if (!gst_multi_fd_sink_handle_client_read (sink, client)) {
        gst_multi_fd_sink_remove_client_link (sink, clients);
        continue;
      }
    }
    if (gst_fdset_fd_can_write (sink->fdset, &client->fd)) {
      /* handle client write */
      if (!gst_multi_fd_sink_handle_client_write (sink, client)) {
        gst_multi_fd_sink_remove_client_link (sink, clients);
        continue;
      }
    }
  }
  CLIENTS_UNLOCK (sink);
}

/* we handle the client communication in another thread so that we do not block
 * the gstreamer thread while we select() on the client fds */
static gpointer
gst_multi_fd_sink_thread (GstMultiFdSink * sink)
{
  while (sink->running) {
    gst_multi_fd_sink_handle_clients (sink);
  }
  return NULL;
}

static GstFlowReturn
gst_multi_fd_sink_render (GstBaseSink * bsink, GstBuffer * buf)
{
  GstMultiFdSink *sink;
  gboolean in_caps;
  GstCaps *bufcaps, *padcaps;

  sink = GST_MULTI_FD_SINK (bsink);

  /* since we check every buffer for streamheader caps, we need to make
   * sure every buffer has caps set */
  bufcaps = gst_buffer_get_caps (buf);
  padcaps = GST_PAD_CAPS (GST_BASE_SINK_PAD (bsink));

  /* make sure we have caps on the pad */
  if (!padcaps) {
    if (!bufcaps) {
      GST_ELEMENT_ERROR (sink, CORE, NEGOTIATION, (NULL),
          ("Received first buffer without caps set"));
      return GST_FLOW_NOT_NEGOTIATED;
    }
  }

  /* stamp the buffer with previous caps if no caps set */
  if (!bufcaps) {
    buf = gst_buffer_make_writable (buf);
    gst_buffer_set_caps (buf, padcaps);
  } else {
    gst_caps_unref (bufcaps);
  }

  /* since we keep this buffer out of the scope of this method */
  gst_buffer_ref (buf);

  g_return_val_if_fail (GST_OBJECT_FLAG_IS_SET (sink, GST_MULTI_FD_SINK_OPEN),
      GST_FLOW_ERROR);

  in_caps = GST_BUFFER_FLAG_IS_SET (buf, GST_BUFFER_FLAG_IN_CAPS);

  GST_LOG_OBJECT (sink, "received buffer %p, in_caps: %d", buf, in_caps);

  /* if we get IN_CAPS buffers, but the previous buffer was not IN_CAPS,
   * it means we're getting new streamheader buffers, and we should clear
   * the old ones */
  if (in_caps && sink->previous_buffer_in_caps == FALSE) {
    GST_DEBUG_OBJECT (sink,
        "receiving new IN_CAPS buffers, clearing old streamheader");
    g_slist_foreach (sink->streamheader, (GFunc) gst_mini_object_unref, NULL);
    g_slist_free (sink->streamheader);
    sink->streamheader = NULL;
  }

  /* save the current in_caps */
  sink->previous_buffer_in_caps = in_caps;

  /* if the incoming buffer is marked as IN CAPS, then we assume for now
   * it's a streamheader that needs to be sent to each new client, so we
   * put it on our internal list of streamheader buffers.
   * FIXME: we could check if the buffer's contents are in fact part of the
   * current streamheader.
   *
   * We don't send the buffer to the client, since streamheaders are sent
   * separately when necessary. */
  if (in_caps) {
    GST_DEBUG_OBJECT (sink,
        "appending IN_CAPS buffer with length %d to streamheader",
        GST_BUFFER_SIZE (buf));
    sink->streamheader = g_slist_append (sink->streamheader, buf);
  } else {
    /* queue the buffer, this is a regular data buffer. */
    gst_multi_fd_sink_queue_buffer (sink, buf);

    sink->bytes_to_serve += GST_BUFFER_SIZE (buf);
  }

  return GST_FLOW_OK;
}

static void
gst_multi_fd_sink_set_property (GObject * object, guint prop_id,
    const GValue * value, GParamSpec * pspec)
{
  GstMultiFdSink *multifdsink;

  multifdsink = GST_MULTI_FD_SINK (object);

  switch (prop_id) {
    case PROP_PROTOCOL:
      multifdsink->protocol = g_value_get_enum (value);
      break;
    case PROP_MODE:
      multifdsink->mode = g_value_get_enum (value);
      break;
    case PROP_BUFFERS_MAX:
      multifdsink->units_max = g_value_get_int (value);
      break;
    case PROP_BUFFERS_SOFT_MAX:
      multifdsink->units_soft_max = g_value_get_int (value);
      break;
    case PROP_TIME_MIN:
      multifdsink->time_min = g_value_get_int64 (value);
      break;
    case PROP_BYTES_MIN:
      multifdsink->bytes_min = g_value_get_int (value);
      break;
    case PROP_BUFFERS_MIN:
      multifdsink->buffers_min = g_value_get_int (value);
      break;
    case PROP_UNIT_TYPE:
      multifdsink->unit_type = g_value_get_enum (value);
      break;
    case PROP_UNITS_MAX:
      multifdsink->units_max = g_value_get_int (value);
      break;
    case PROP_UNITS_SOFT_MAX:
      multifdsink->units_soft_max = g_value_get_int (value);
      break;
    case PROP_RECOVER_POLICY:
      multifdsink->recover_policy = g_value_get_enum (value);
      break;
    case PROP_TIMEOUT:
      multifdsink->timeout = g_value_get_uint64 (value);
      break;
    case PROP_SYNC_METHOD:
      multifdsink->def_sync_method = g_value_get_enum (value);
      break;
    case PROP_BURST_UNIT:
      multifdsink->def_burst_unit = g_value_get_enum (value);
      break;
    case PROP_BURST_VALUE:
      multifdsink->def_burst_value = g_value_get_uint64 (value);
      break;

    default:
      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
      break;
  }
}

static void
gst_multi_fd_sink_get_property (GObject * object, guint prop_id, GValue * value,
    GParamSpec * pspec)
{
  GstMultiFdSink *multifdsink;

  multifdsink = GST_MULTI_FD_SINK (object);

  switch (prop_id) {
    case PROP_PROTOCOL:
      g_value_set_enum (value, multifdsink->protocol);
      break;
    case PROP_MODE:
      g_value_set_enum (value, multifdsink->mode);
      break;
    case PROP_BUFFERS_MAX:
      g_value_set_int (value, multifdsink->units_max);
      break;
    case PROP_BUFFERS_SOFT_MAX:
      g_value_set_int (value, multifdsink->units_soft_max);
      break;
    case PROP_TIME_MIN:
      g_value_set_int64 (value, multifdsink->time_min);
      break;
    case PROP_BYTES_MIN:
      g_value_set_int (value, multifdsink->bytes_min);
      break;
    case PROP_BUFFERS_MIN:
      g_value_set_int (value, multifdsink->buffers_min);
      break;
    case PROP_BUFFERS_QUEUED:
      g_value_set_uint (value, multifdsink->buffers_queued);
      break;
    case PROP_BYTES_QUEUED:
      g_value_set_uint (value, multifdsink->bytes_queued);
      break;
    case PROP_TIME_QUEUED:
      g_value_set_uint64 (value, multifdsink->time_queued);
      break;
    case PROP_UNIT_TYPE:
      g_value_set_enum (value, multifdsink->unit_type);
      break;
    case PROP_UNITS_MAX:
      g_value_set_int (value, multifdsink->units_max);
      break;
    case PROP_UNITS_SOFT_MAX:
      g_value_set_int (value, multifdsink->units_soft_max);
      break;
    case PROP_RECOVER_POLICY:
      g_value_set_enum (value, multifdsink->recover_policy);
      break;
    case PROP_TIMEOUT:
      g_value_set_uint64 (value, multifdsink->timeout);
      break;
    case PROP_SYNC_METHOD:
      g_value_set_enum (value, multifdsink->def_sync_method);
      break;
    case PROP_BYTES_TO_SERVE:
      g_value_set_uint64 (value, multifdsink->bytes_to_serve);
      break;
    case PROP_BYTES_SERVED:
      g_value_set_uint64 (value, multifdsink->bytes_served);
      break;
    case PROP_BURST_UNIT:
      g_value_set_enum (value, multifdsink->def_burst_unit);
      break;
    case PROP_BURST_VALUE:
      g_value_set_uint64 (value, multifdsink->def_burst_value);
      break;

    default:
      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
      break;
  }
}


/* create a socket for sending to remote machine */
static gboolean
gst_multi_fd_sink_start (GstBaseSink * bsink)
{
  GstMultiFdSinkClass *fclass;
  int control_socket[2];
  GstMultiFdSink *this;

  if (GST_OBJECT_FLAG_IS_SET (bsink, GST_MULTI_FD_SINK_OPEN))
    return TRUE;

  this = GST_MULTI_FD_SINK (bsink);
  fclass = GST_MULTI_FD_SINK_GET_CLASS (this);

  GST_INFO_OBJECT (this, "starting in mode %d", this->mode);
  this->fdset = gst_fdset_new (this->mode);

  if (socketpair (PF_UNIX, SOCK_STREAM, 0, control_socket) < 0)
    goto socket_pair;

  READ_SOCKET (this).fd = control_socket[0];
  WRITE_SOCKET (this).fd = control_socket[1];

  gst_fdset_add_fd (this->fdset, &READ_SOCKET (this));
  gst_fdset_fd_ctl_read (this->fdset, &READ_SOCKET (this), TRUE);

  fcntl (READ_SOCKET (this).fd, F_SETFL, O_NONBLOCK);
  fcntl (WRITE_SOCKET (this).fd, F_SETFL, O_NONBLOCK);

  this->streamheader = NULL;
  this->bytes_to_serve = 0;
  this->bytes_served = 0;

  if (fclass->init) {
    fclass->init (this);
  }

  this->running = TRUE;
  this->thread = g_thread_create ((GThreadFunc) gst_multi_fd_sink_thread,
      this, TRUE, NULL);

  GST_OBJECT_FLAG_SET (this, GST_MULTI_FD_SINK_OPEN);

  return TRUE;

  /* ERRORS */
socket_pair:
  {
    GST_ELEMENT_ERROR (this, RESOURCE, OPEN_READ_WRITE, (NULL),
        GST_ERROR_SYSTEM);
    return FALSE;
  }
}

static gboolean
multifdsink_hash_remove (gpointer key, gpointer value, gpointer data)
{
  return TRUE;
}

static gboolean
gst_multi_fd_sink_stop (GstBaseSink * bsink)
{
  GstMultiFdSinkClass *fclass;
  GstMultiFdSink *this;
  GstBuffer *buf;
  int i;

  this = GST_MULTI_FD_SINK (bsink);
  fclass = GST_MULTI_FD_SINK_GET_CLASS (this);

  if (!GST_OBJECT_FLAG_IS_SET (bsink, GST_MULTI_FD_SINK_OPEN))
    return TRUE;

  this->running = FALSE;

  SEND_COMMAND (this, CONTROL_STOP);
  if (this->thread) {
    GST_DEBUG_OBJECT (this, "joining thread");
    g_thread_join (this->thread);
    GST_DEBUG_OBJECT (this, "joined thread");
    this->thread = NULL;
  }

  /* free the clients */
  gst_multi_fd_sink_clear (this);

  close (READ_SOCKET (this).fd);
  close (WRITE_SOCKET (this).fd);

  if (this->streamheader) {
    g_slist_foreach (this->streamheader, (GFunc) gst_mini_object_unref, NULL);
    g_slist_free (this->streamheader);
    this->streamheader = NULL;
  }

  if (fclass->close)
    fclass->close (this);

  if (this->fdset) {
    gst_fdset_remove_fd (this->fdset, &READ_SOCKET (this));
    gst_fdset_free (this->fdset);
    this->fdset = NULL;
  }
  g_hash_table_foreach_remove (this->fd_hash, multifdsink_hash_remove, this);

  /* remove all queued buffers */
  if (this->bufqueue) {
    GST_DEBUG_OBJECT (this, "Emptying bufqueue with %d buffers",
        this->bufqueue->len);
    for (i = this->bufqueue->len - 1; i >= 0; --i) {
      buf = g_array_index (this->bufqueue, GstBuffer *, i);
      GST_LOG_OBJECT (this, "Removing buffer %p (%d) with refcount %d", buf, i,
          GST_MINI_OBJECT_REFCOUNT (buf));
      gst_buffer_unref (buf);
      this->bufqueue = g_array_remove_index (this->bufqueue, i);
    }
    /* freeing the array is done in _finalize */
  }
  GST_OBJECT_FLAG_UNSET (this, GST_MULTI_FD_SINK_OPEN);

  return TRUE;
}

static GstStateChangeReturn
gst_multi_fd_sink_change_state (GstElement * element, GstStateChange transition)
{
  GstMultiFdSink *sink;
  GstStateChangeReturn ret;

  sink = GST_MULTI_FD_SINK (element);

  /* we disallow changing the state from the streaming thread */
  if (g_thread_self () == sink->thread)
    return GST_STATE_CHANGE_FAILURE;


  switch (transition) {
    case GST_STATE_CHANGE_NULL_TO_READY:
      if (!gst_multi_fd_sink_start (GST_BASE_SINK (sink)))
        goto start_failed;
      break;
    case GST_STATE_CHANGE_READY_TO_PAUSED:
      break;
    case GST_STATE_CHANGE_PAUSED_TO_PLAYING:
      break;
    default:
      break;
  }

  ret = GST_ELEMENT_CLASS (parent_class)->change_state (element, transition);

  switch (transition) {
    case GST_STATE_CHANGE_PLAYING_TO_PAUSED:
      break;
    case GST_STATE_CHANGE_PAUSED_TO_READY:
      break;
    case GST_STATE_CHANGE_READY_TO_NULL:
      gst_multi_fd_sink_stop (GST_BASE_SINK (sink));
      break;
    default:
      break;
  }
  return ret;

  /* ERRORS */
start_failed:
  {
    /* error message was posted */
    return GST_STATE_CHANGE_FAILURE;
  }
}