* Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
* Boston, MA 02110-1301, USA.
*/
+/**
+ * SECTION:element-latencytracer
+ * @short_description: log processing latency stats
+ *
+ * A tracing module that determines src-to-sink latencies by injecting custom
+ * events at sources and process them at sinks. This elements supports tracing
+ * the entire pipeline latency and per element latency. By default, only
+ * pipeline latency is traced. The 'flags' parameter can be used to enabled
+ * element tracing.
+ *
+ * ```
+ * GST_TRACERS="latency(flags=pipeline+element)" GST_DEBUG=GST_TRACER:7 ./...
+ * ```
+ */
+/* TODO(ensonic): if there are two sources feeding into a mixer/muxer and later
+ * we fan-out with tee and have two sinks, each sink would get all two events,
+ * the later event would overwrite the former. Unfortunately when the buffer
+ * arrives on the sink we don't know to which event it correlates. Better would
+ * be to use the buffer meta in 1.0 instead of the event. Or we track a min/max
+ * latency.
+ */
#ifdef HAVE_CONFIG_H
# include "config.h"
_do_init);
static GQuark latency_probe_id;
+static GQuark sub_latency_probe_id;
static GQuark latency_probe_pad;
static GQuark latency_probe_ts;
+static GQuark drop_sub_latency_quark;
+
+static GstTracerRecord *tr_latency;
+static GstTracerRecord *tr_element_latency;
/* data helpers */
/*
- * Get the element/bin owning the pad.
+ * Get the element/bin owning the pad.
*
* in: a normal pad
* out: the element
return GST_ELEMENT_CAST (parent);
}
-/* tracer class */
-
-static void gst_latency_tracer_invoke (GstTracer * obj, GstTracerHookId id,
- GstTracerMessageId mid, va_list var_args);
+/* hooks */
static void
-gst_latency_tracer_class_init (GstLatencyTracerClass * klass)
+log_latency (const GstStructure * data, GstPad * sink_pad, guint64 sink_ts)
{
- GstTracerClass *gst_tracer_class = GST_TRACER_CLASS (klass);
+ guint64 src_ts;
+ const char *src;
+ const GValue *value;
+ gchar *sink;
- gst_tracer_class->invoke = gst_latency_tracer_invoke;
+ value = gst_structure_id_get_value (data, latency_probe_ts);
+ src_ts = g_value_get_uint64 (value);
- latency_probe_id = g_quark_from_static_string ("latency_probe.id");
- latency_probe_pad = g_quark_from_static_string ("latency_probe.pad");
- latency_probe_ts = g_quark_from_static_string ("latency_probe.ts");
+ value = gst_structure_id_get_value (data, latency_probe_pad);
+ src = g_value_get_string (value);
+
+ sink = g_strdup_printf ("%s_%s",
+ GST_DEBUG_PAD_NAME (GST_PAD_PEER (sink_pad)));
+ gst_tracer_record_log (tr_latency, src, sink,
+ GST_CLOCK_DIFF (src_ts, sink_ts), sink_ts);
+ g_free (sink);
}
static void
-gst_latency_tracer_init (GstLatencyTracer * self)
+log_element_latency (const GstStructure * data, GstPad * pad, guint64 sink_ts)
{
- g_object_set (self, "mask", GST_TRACER_HOOK_BUFFERS | GST_TRACER_HOOK_EVENTS,
- NULL);
-}
+ guint64 src_ts;
+ gchar *pad_name;
+ const GValue *value;
-/* hooks */
+ pad_name = g_strdup_printf ("%s_%s", GST_DEBUG_PAD_NAME (pad));
+
+ /* TODO filtering */
+
+ value = gst_structure_id_get_value (data, latency_probe_ts);
+ src_ts = g_value_get_uint64 (value);
+
+ gst_tracer_record_log (tr_element_latency, pad_name,
+ GST_CLOCK_DIFF (src_ts, sink_ts), sink_ts);
+
+ g_free (pad_name);
+}
static void
send_latency_probe (GstLatencyTracer * self, GstElement * parent, GstPad * pad,
guint64 ts)
{
+ GstPad *peer_pad = GST_PAD_PEER (pad);
+
+ /* allow for non-parented pads to send latency probes as used in e.g.
+ * rtspsrc for TCP connections */
+ if (peer_pad && (!parent || (!GST_IS_BIN (parent)))) {
+ gchar *pad_name;
+ GstEvent *latency_probe;
+
+ if (self->flags & GST_LATENCY_TRACER_FLAG_PIPELINE &&
+ GST_OBJECT_FLAG_IS_SET (parent, GST_ELEMENT_FLAG_SOURCE)) {
+ pad_name = g_strdup_printf ("%s_%s", GST_DEBUG_PAD_NAME (pad));
+
+ GST_DEBUG ("%s: Sending latency event", pad_name);
+
+ latency_probe = gst_event_new_custom (GST_EVENT_CUSTOM_DOWNSTREAM,
+ gst_structure_new_id (latency_probe_id,
+ latency_probe_pad, G_TYPE_STRING, pad_name,
+ latency_probe_ts, G_TYPE_UINT64, ts, NULL));
+ g_free (pad_name);
+ gst_pad_push_event (pad, latency_probe);
+ }
+
+ if (self->flags & GST_LATENCY_TRACER_FLAG_ELEMENT) {
+ GST_DEBUG ("%s_%s: Sending sub-latency event", GST_DEBUG_PAD_NAME (pad));
+
+ pad_name = g_strdup_printf ("%s_%s", GST_DEBUG_PAD_NAME (peer_pad));
+ latency_probe = gst_event_new_custom (GST_EVENT_CUSTOM_DOWNSTREAM,
+ gst_structure_new_id (sub_latency_probe_id,
+ latency_probe_pad, G_TYPE_STRING, pad_name,
+ latency_probe_ts, G_TYPE_UINT64, ts, NULL));
+ gst_pad_push_event (pad, latency_probe);
+ g_free (pad_name);
+ }
+ }
+}
+
+static void
+calculate_latency (GstElement * parent, GstPad * pad, guint64 ts)
+{
+ GstElement *peer_parent = get_real_pad_parent (GST_PAD_PEER (pad));
+
if (parent && (!GST_IS_BIN (parent)) &&
- GST_OBJECT_FLAG_IS_SET (parent, GST_ELEMENT_FLAG_SOURCE)) {
- GstEvent *latency_probe = gst_event_new_custom (GST_EVENT_CUSTOM_DOWNSTREAM,
- gst_structure_new_id (latency_probe_id,
- latency_probe_pad, GST_TYPE_PAD, pad,
- latency_probe_ts, G_TYPE_UINT64, ts,
- NULL));
- gst_pad_push_event (pad, latency_probe);
+ (!GST_OBJECT_FLAG_IS_SET (parent, GST_ELEMENT_FLAG_SOURCE))) {
+ GstEvent *ev;
+
+ /* FIXME unsafe use of peer */
+ if (GST_OBJECT_FLAG_IS_SET (peer_parent, GST_ELEMENT_FLAG_SINK)) {
+ GST_DEBUG ("%s_%s: Should log full lantency now",
+ GST_DEBUG_PAD_NAME (pad));
+ ev = g_object_get_qdata ((GObject *) pad, latency_probe_id);
+ if (ev) {
+ g_object_set_qdata ((GObject *) pad, latency_probe_id, NULL);
+ log_latency (gst_event_get_structure (ev), pad, ts);
+ gst_event_unref (ev);
+ }
+ }
+
+ GST_DEBUG ("%s_%s: Should log sub lantency now", GST_DEBUG_PAD_NAME (pad));
+ ev = g_object_get_qdata ((GObject *) pad, sub_latency_probe_id);
+ if (ev) {
+ g_object_set_qdata ((GObject *) pad, sub_latency_probe_id, NULL);
+ log_element_latency (gst_event_get_structure (ev), pad, ts);
+ gst_event_unref (ev);
+ }
}
}
static void
-do_push_buffer_pre (GstLatencyTracer * self, va_list var_args)
+do_push_buffer_pre (GstTracer * tracer, guint64 ts, GstPad * pad)
{
- guint64 ts = va_arg (var_args, guint64);
- GstPad *pad = va_arg (var_args, GstPad *);
+ GstLatencyTracer *self = (GstLatencyTracer *) tracer;
GstElement *parent = get_real_pad_parent (pad);
send_latency_probe (self, parent, pad, ts);
+ calculate_latency (parent, pad, ts);
}
static void
-do_pull_buffer_pre (GstLatencyTracer * self, va_list var_args)
+do_pull_range_pre (GstTracer * tracer, guint64 ts, GstPad * pad)
{
- guint64 ts = va_arg (var_args, guint64);
- GstPad *pad = va_arg (var_args, GstPad *);
+ GstLatencyTracer *self = (GstLatencyTracer *) tracer;
GstPad *peer_pad = GST_PAD_PEER (pad);
GstElement *parent = get_real_pad_parent (peer_pad);
}
static void
-do_push_event_pre (GstLatencyTracer * self, va_list var_args)
+do_pull_range_post (GstTracer * self, guint64 ts, GstPad * pad)
+{
+ GstElement *parent = get_real_pad_parent (pad);
+
+ calculate_latency (parent, pad, ts);
+}
+
+static GstPadProbeReturn
+do_drop_sub_latency_event (GstPad * pad, GstPadProbeInfo * info,
+ gpointer user_data)
+{
+ GstEvent *ev = info->data;
+ GstPadProbeReturn ret = GST_PAD_PROBE_OK;
+
+ if (GST_EVENT_TYPE (ev) == GST_EVENT_CUSTOM_DOWNSTREAM) {
+ const GstStructure *data = gst_event_get_structure (ev);
+
+ if (gst_structure_get_name_id (data) == sub_latency_probe_id) {
+ const GValue *value;
+ /* FIXME unsafe peer pad usage */
+ gchar *pad_name = g_strdup_printf ("%s_%s",
+ GST_DEBUG_PAD_NAME (GST_PAD_PEER (pad)));
+
+ value = gst_structure_id_get_value (data, latency_probe_pad);
+ if (!g_str_equal (g_value_get_string (value), pad_name)) {
+ GST_DEBUG ("%s: Dropping sub-latency event", pad_name);
+ ret = GST_PAD_PROBE_DROP;
+ }
+
+ g_free (pad_name);
+ }
+ }
+
+ return ret;
+}
+
+static void
+do_push_event_pre (GstTracer * self, guint64 ts, GstPad * pad, GstEvent * ev)
{
- guint64 ts = va_arg (var_args, guint64);
- GstPad *pad = va_arg (var_args, GstPad *);
- GstEvent *ev = va_arg (var_args, GstEvent *);
+ GstElement *parent = get_real_pad_parent (pad);
GstPad *peer_pad = GST_PAD_PEER (pad);
- GstElement *parent = get_real_pad_parent (peer_pad);
+ GstElement *peer_parent = get_real_pad_parent (peer_pad);
if (parent && (!GST_IS_BIN (parent)) &&
- GST_OBJECT_FLAG_IS_SET (parent, GST_ELEMENT_FLAG_SINK)) {
- if (GST_EVENT_TYPE (ev) == GST_EVENT_CUSTOM_DOWNSTREAM) {
- const GstStructure *data = gst_event_get_structure (ev);
-
- if (gst_structure_get_name_id (data) == latency_probe_id) {
- GstPad *origin_pad;
- guint64 origin_ts;
- gchar *from, *to;
-
- /* TODO(ensonic): we'd like to do this when actually rendering */
- gst_structure_id_get (data,
- latency_probe_pad, GST_TYPE_PAD, &origin_pad,
- latency_probe_ts, G_TYPE_UINT64, &origin_ts, NULL);
-
- from = g_strdup_printf ("%s_%s", GST_DEBUG_PAD_NAME (origin_pad));
- to = g_strdup_printf ("%s_%s", GST_DEBUG_PAD_NAME (peer_pad));
-
- /* TODO(ensonic): report format is still unstable */
- gst_tracer_log_trace (gst_structure_new ("latency",
- "from", G_TYPE_STRING, from,
- "to", G_TYPE_STRING, to,
- "time", G_TYPE_UINT64, GST_CLOCK_DIFF (origin_ts, ts), NULL));
- g_free (from);
- g_free (to);
+ (!GST_OBJECT_FLAG_IS_SET (parent, GST_ELEMENT_FLAG_SOURCE)) &&
+ GST_EVENT_TYPE (ev) == GST_EVENT_CUSTOM_DOWNSTREAM) {
+ const GstStructure *data = gst_event_get_structure (ev);
+
+ /* if not set yet, add a pad probe that prevents sub-latency event from
+ * flowing further */
+ if (gst_structure_get_name_id (data) == latency_probe_id) {
+
+ if (!g_object_get_qdata ((GObject *) pad, drop_sub_latency_quark)) {
+ GST_DEBUG ("%s_%s: Adding pad probe to drop sub-latency event",
+ GST_DEBUG_PAD_NAME (pad));
+ gst_pad_add_probe (pad, GST_PAD_PROBE_TYPE_EVENT_DOWNSTREAM,
+ do_drop_sub_latency_event, NULL, NULL);
+ g_object_set_qdata ((GObject *) pad, drop_sub_latency_quark,
+ (gpointer) 1);
+ }
+
+ /* FIXME unsafe peer access */
+ if (GST_OBJECT_FLAG_IS_SET (peer_parent, GST_ELEMENT_FLAG_SINK)) {
+ GST_DEBUG ("%s_%s: Storing latency event", GST_DEBUG_PAD_NAME (pad));
+
+ /* store event so that we can calculate latency when the buffer that
+ * follows has been processed */
+ if (!g_object_get_qdata ((GObject *) pad, latency_probe_id))
+ g_object_set_qdata ((GObject *) pad, latency_probe_id,
+ gst_event_ref (ev));
}
}
+
+ if (gst_structure_get_name_id (data) == sub_latency_probe_id) {
+ const GValue *value;
+ gchar *pad_name = g_strdup_printf ("%s_%s",
+ GST_DEBUG_PAD_NAME (peer_pad));
+
+ value = gst_structure_id_get_value (data, latency_probe_pad);
+
+ if (!g_str_equal (g_value_get_string (value), pad_name)) {
+ GST_DEBUG ("%s: Storing sub-latency event", pad_name);
+ if (!g_object_get_qdata ((GObject *) pad, sub_latency_probe_id))
+ g_object_set_qdata ((GObject *) pad, sub_latency_probe_id,
+ gst_event_ref (ev));
+ }
+
+ g_free (pad_name);
+ }
}
}
+/* tracer class */
+
static void
-gst_latency_tracer_invoke (GstTracer * obj, GstTracerHookId hid,
- GstTracerMessageId mid, va_list var_args)
+gst_latency_tracer_constructed (GObject * object)
{
- GstLatencyTracer *self = GST_LATENCY_TRACER_CAST (obj);
-
- switch (mid) {
- case GST_TRACER_MESSAGE_ID_PAD_PUSH_PRE:
- case GST_TRACER_MESSAGE_ID_PAD_PUSH_LIST_PRE:
- do_push_buffer_pre (self, var_args);
- break;
- case GST_TRACER_MESSAGE_ID_PAD_PULL_RANGE_PRE:
- do_pull_buffer_pre (self, var_args);
- break;
- case GST_TRACER_MESSAGE_ID_PAD_PUSH_EVENT_PRE:
- do_push_event_pre (self, var_args);
- break;
- default:
- break;
+ GstLatencyTracer *self = GST_LATENCY_TRACER (object);
+ gchar *params, *tmp;
+ GstStructure *params_struct = NULL;
+
+ g_object_get (self, "params", ¶ms, NULL);
+
+ if (!params)
+ return;
+
+ tmp = g_strdup_printf ("latency,%s", params);
+ params_struct = gst_structure_from_string (tmp, NULL);
+ g_free (tmp);
+
+ /* Read the flags if available */
+ if (params_struct) {
+ const gchar *flags = gst_structure_get_string (params_struct, "flags");
+
+ self->flags = 0;
+
+ if (flags) {
+ GStrv split = g_strsplit (flags, "+", -1);
+ gint i;
+
+ for (i = 0; split[i]; i++) {
+ if (g_str_equal (split[i], "pipeline"))
+ self->flags |= GST_LATENCY_TRACER_FLAG_PIPELINE;
+ else if (g_str_equal (split[i], "element"))
+ self->flags |= GST_LATENCY_TRACER_FLAG_ELEMENT;
+ else
+ GST_WARNING ("Invalid latency tracer flags %s", split[i]);
+ }
+
+ g_strfreev (split);
+ }
}
+
+ g_free (params);
+}
+
+static void
+gst_latency_tracer_class_init (GstLatencyTracerClass * klass)
+{
+ GObjectClass *gobject_class = G_OBJECT_CLASS (klass);
+
+ gobject_class->constructed = gst_latency_tracer_constructed;
+
+ latency_probe_id = g_quark_from_static_string ("latency_probe.id");
+ sub_latency_probe_id = g_quark_from_static_string ("sub_latency_probe.id");
+ latency_probe_pad = g_quark_from_static_string ("latency_probe.pad");
+ latency_probe_ts = g_quark_from_static_string ("latency_probe.ts");
+ drop_sub_latency_quark =
+ g_quark_from_static_string ("drop_sub_latency.quark");
+
+ /* announce trace formats */
+ /* *INDENT-OFF* */
+ tr_latency = gst_tracer_record_new ("latency.class",
+ "src", GST_TYPE_STRUCTURE, gst_structure_new ("scope",
+ "type", G_TYPE_GTYPE, G_TYPE_STRING,
+ "related-to", GST_TYPE_TRACER_VALUE_SCOPE, GST_TRACER_VALUE_SCOPE_PAD,
+ NULL),
+ "sink", GST_TYPE_STRUCTURE, gst_structure_new ("scope",
+ "type", G_TYPE_GTYPE, G_TYPE_STRING,
+ "related-to", GST_TYPE_TRACER_VALUE_SCOPE, GST_TRACER_VALUE_SCOPE_PAD,
+ NULL),
+ "time", GST_TYPE_STRUCTURE, gst_structure_new ("value",
+ "type", G_TYPE_GTYPE, G_TYPE_UINT64,
+ "description", G_TYPE_STRING,
+ "time it took for the buffer to go from src to sink ns",
+ "min", G_TYPE_UINT64, G_GUINT64_CONSTANT (0),
+ "max", G_TYPE_UINT64, G_MAXUINT64,
+ NULL),
+ "ts", GST_TYPE_STRUCTURE, gst_structure_new ("value",
+ "type", G_TYPE_GTYPE, G_TYPE_UINT64,
+ "description", G_TYPE_STRING, "ts when the latency has been logged",
+ "min", G_TYPE_UINT64, G_GUINT64_CONSTANT (0),
+ "max", G_TYPE_UINT64, G_MAXUINT64,
+ NULL),
+ NULL);
+
+ tr_element_latency = gst_tracer_record_new ("element-latency.class",
+ "src", GST_TYPE_STRUCTURE, gst_structure_new ("scope",
+ "type", G_TYPE_GTYPE, G_TYPE_STRING,
+ "related-to", GST_TYPE_TRACER_VALUE_SCOPE, GST_TRACER_VALUE_SCOPE_PAD,
+ NULL),
+ "time", GST_TYPE_STRUCTURE, gst_structure_new ("value",
+ "type", G_TYPE_GTYPE, G_TYPE_UINT64,
+ "description", G_TYPE_STRING,
+ "time it took for the buffer to go from src to sink ns",
+ "min", G_TYPE_UINT64, G_GUINT64_CONSTANT (0),
+ "max", G_TYPE_UINT64, G_MAXUINT64,
+ NULL),
+ "ts", GST_TYPE_STRUCTURE, gst_structure_new ("value",
+ "type", G_TYPE_GTYPE, G_TYPE_UINT64,
+ "description", G_TYPE_STRING, "ts when the latency has been logged",
+ "min", G_TYPE_UINT64, G_GUINT64_CONSTANT (0),
+ "max", G_TYPE_UINT64, G_MAXUINT64,
+ NULL),
+ NULL);
+ /* *INDENT-ON* */
+}
+
+static void
+gst_latency_tracer_init (GstLatencyTracer * self)
+{
+ GstTracer *tracer = GST_TRACER (self);
+
+ /* only trace pipeline latency by default */
+ self->flags = GST_LATENCY_TRACER_FLAG_PIPELINE;
+
+ /* in push mode, pre/post will be called before/after the peer chain
+ * function has been called. For this reaosn, we only use -pre to avoid
+ * accounting for the processing time of the peer element (the sink) */
+ gst_tracing_register_hook (tracer, "pad-push-pre",
+ G_CALLBACK (do_push_buffer_pre));
+ gst_tracing_register_hook (tracer, "pad-push-list-pre",
+ G_CALLBACK (do_push_buffer_pre));
+
+ /* while in pull mode, pre/post will happen before and after the upstream
+ * pull_range call is made, so it already only account for the upstream
+ * processing time. As a side effect, in pull mode, we can measure the
+ * source processing latency, while in push mode, we can't */
+ gst_tracing_register_hook (tracer, "pad-pull-range-pre",
+ G_CALLBACK (do_pull_range_pre));
+ gst_tracing_register_hook (tracer, "pad-pull-range-post",
+ G_CALLBACK (do_pull_range_post));
+
+ gst_tracing_register_hook (tracer, "pad-push-event-pre",
+ G_CALLBACK (do_push_event_pre));
}
projects / platform / upstream / gstreamer.git / blobdiff