gstreamer/subprojects/gst-plugins-bad/ext/dash/gstdashdemux.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

4096 lines
144 KiB
C
Raw Normal View History

2013-05-08 14:13:32 +00:00
/*
* DASH demux plugin for GStreamer
*
* gstdashdemux.c
*
* Copyright (C) 2012 Orange
2014-08-26 19:45:46 +00:00
*
2013-05-08 14:13:32 +00:00
* Authors:
* David Corvoysier <david.corvoysier@orange.com>
* Hamid Zakari <hamid.zakari@gmail.com>
*
* Copyright (C) 2013 Smart TV Alliance
* Author: Thiago Sousa Santos <thiago.sousa.santos@collabora.com>, Collabora Ltd.
*
2013-05-08 14:13:32 +00:00
* 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 2.1 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
2012-09-28 10:55:27 +00:00
* License along with this library (COPYING); if not, write to the
2013-05-08 14:13:32 +00:00
* Free Software Foundation, Inc., 59 Temple Place - Suite 330,
* Boston, MA 02111-1307, USA.
*/
/**
* SECTION:element-dashdemux
* @title: dashdemux
2013-05-08 14:13:32 +00:00
*
* DASH demuxer element.
* ## Example launch line
2013-05-08 14:13:32 +00:00
* |[
* gst-launch-1.0 playbin uri="http://www-itec.uni-klu.ac.at/ftp/datasets/mmsys12/RedBullPlayStreets/redbull_4s/RedBullPlayStreets_4s_isoffmain_DIS_23009_1_v_2_1c2_2011_08_30.mpd"
2013-05-08 14:13:32 +00:00
* ]|
*/
/* Implementation notes:
2014-08-26 19:45:46 +00:00
*
* The following section describes how dashdemux works internally.
2014-08-26 19:45:46 +00:00
*
* Introduction:
2014-08-26 19:45:46 +00:00
*
* dashdemux is a "fake" demux, as unlike traditional demux elements, it
* doesn't split data streams contained in an envelope to expose them to
* downstream decoding elements.
2014-08-26 19:45:46 +00:00
*
* Instead, it parses an XML file called a manifest to identify a set of
* individual stream fragments it needs to fetch and expose to the actual
2014-08-26 19:45:46 +00:00
* demux elements that will handle them (this behavior is sometimes
* referred as the "demux after a demux" scenario).
2014-08-26 19:45:46 +00:00
*
* For a given section of content, several representations corresponding
* to different bitrates may be available: dashdemux will select the most
2014-08-26 19:45:46 +00:00
* appropriate representation based on local conditions (typically the
* available bandwidth and the amount of buffering available, capped by
2014-08-26 19:45:46 +00:00
* a maximum allowed bitrate).
*
* The representation selection algorithm can be configured using
* specific properties: max bitrate, min/max buffering, bandwidth ratio.
2014-08-26 19:45:46 +00:00
*
*
* General Design:
2014-08-26 19:45:46 +00:00
*
* dashdemux has a single sink pad that accepts the data corresponding
* to the manifest, typically fetched from an HTTP or file source.
2014-08-26 19:45:46 +00:00
*
* dashdemux exposes the streams it recreates based on the fragments it
* fetches through dedicated src pads corresponding to the caps of the
* fragments container (ISOBMFF/MP4 or MPEG2TS).
2014-08-26 19:45:46 +00:00
*
* During playback, new representations will typically be exposed as a
* new set of pads (see 'Switching between representations' below).
2014-08-26 19:45:46 +00:00
*
* Fragments downloading is performed using a dedicated task that fills
* an internal queue. Another task is in charge of popping fragments
* from the queue and pushing them downstream.
2014-08-26 19:45:46 +00:00
*
* Switching between representations:
2014-08-26 19:45:46 +00:00
*
* Decodebin supports scenarios allowing to seamlessly switch from one
* stream to another inside the same "decoding chain".
2014-08-26 19:45:46 +00:00
*
* To achieve that, it combines the elements it autoplugged in chains
* and groups, allowing only one decoding group to be active at a given
* time for a given chain.
*
2014-08-26 19:45:46 +00:00
* A chain can signal decodebin that it is complete by sending a
* no-more-pads event, but even after that new pads can be added to
* create new subgroups, providing that a new no-more-pads event is sent.
*
* We take advantage of that to dynamically create a new decoding group
* in order to select a different representation during playback.
*
* Typically, assuming that each fragment contains both audio and video,
* the following tree would be created:
2014-08-26 19:45:46 +00:00
*
* chain "DASH Demux"
* |_ group "Representation set 1"
* | |_ chain "Qt Demux 0"
* | |_ group "Stream 0"
* | |_ chain "H264"
* | |_ chain "AAC"
* |_ group "Representation set 2"
* |_ chain "Qt Demux 1"
* |_ group "Stream 1"
* |_ chain "H264"
* |_ chain "AAC"
*
* Or, if audio and video are contained in separate fragments:
*
* chain "DASH Demux"
* |_ group "Representation set 1"
* | |_ chain "Qt Demux 0"
* | | |_ group "Stream 0"
* | | |_ chain "H264"
* | |_ chain "Qt Demux 1"
* | |_ group "Stream 1"
2014-08-26 19:45:46 +00:00
* | |_ chain "AAC"
* |_ group "Representation set 2"
* |_ chain "Qt Demux 3"
* | |_ group "Stream 2"
* | |_ chain "H264"
* |_ chain "Qt Demux 4"
* |_ group "Stream 3"
2014-08-26 19:45:46 +00:00
* |_ chain "AAC"
*
* In both cases, when switching from Set 1 to Set 2 an EOS is sent on
* each end pad corresponding to Rep 0, triggering the "drain" state to
* propagate upstream.
* Once both EOS have been processed, the "Set 1" group is completely
* drained, and decodebin2 will switch to the "Set 2" group.
2014-08-26 19:45:46 +00:00
*
* Note: nothing can be pushed to the new decoding group before the
* old one has been drained, which means that in order to be able to
* adapt quickly to bandwidth changes, we will not be able to rely
* on downstream buffering, and will instead manage an internal queue.
2014-08-26 19:45:46 +00:00
*
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
*
* Keyframe trick-mode implementation:
*
* When requested (with GST_SEEK_FLAG_TRICKMODE_KEY_UNIT) and if the format
* is supported (ISOBMFF profiles), dashdemux can download only keyframes
* in order to provide fast forward/reverse playback without exceeding the
2019-09-02 19:08:44 +00:00
* available bandwidth/cpu/memory usage.
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
*
* This is done in two parts:
* 1) Parsing ISOBMFF atoms to detect the location of keyframes and only
* download/push those.
* 2) Deciding what the ideal next keyframe to download is in order to
* provide as many keyframes as possible without rebuffering.
*
* * Keyframe-only downloads:
*
* For each beginning of fragment, the fragment header will be parsed in
* gst_dash_demux_parse_isobmff() and then the information (offset, pts...)
* of each keyframe will be stored in moof_sync_samples.
*
* gst_dash_demux_stream_update_fragment_info() will specify the range
* start and end of the current keyframe, which will cause GstAdaptiveDemux
* to do a new upstream range request.
*
* When advancing, if there are still some keyframes in the current
* fragment, gst_dash_demux_stream_advance_fragment() will call
* gst_dash_demux_stream_advance_sync_sample() which decides what the next
* keyframe to get will be (it can be in reverse order for example, or
* might not be the *next* keyframe but one further as explained below).
*
* If no more keyframes are available in the current fragment, dash will
* advance to the next fragment (just like in the normal case) or to a
* fragment much further away (as explained below).
*
*
* * Deciding the optimal "next" keyframe/fragment to download:
*
* The main reason for doing keyframe-only downloads is for trick-modes
* (i.e. being able to do fast reverse/forward playback with limited
2019-09-02 19:08:44 +00:00
* bandwidth/cpu/memory).
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
*
* Downloading all keyframes might not be the optimal solution, especially
* at high playback rates, since the time taken to download the keyframe
* might exceed the available running time between two displayed frames
* (i.e. all frames would end up arriving late). This would cause severe
* rebuffering.
*
* Note: The values specified below can be in either the segment running
* time or in absolute values. Where position values need to be converted
* to segment running time the "running_time(val)" notation is used, and
* where running time need ot be converted to segment poisition the
* "position(val)" notation is used.
*
* The goal instead is to be able to download/display as many frames as
* possible for a given playback rate. For that the implementation will
* take into account:
* * The requested playback rate and segment
* * The average time to request and download a keyframe (in running time)
* * The current position of dashdemux in the stream
* * The current downstream (i.e. sink) position (in running time)
*
* To reach this goal we consider that there is some amount of buffering
* (in time) between dashdemux and the display sink. While we do not know
* the exact amount of buffering available, a safe and reasonable assertion
* is that there is at least a second (in running time).
*
* The average time to request and fully download a keyframe (with or
* without fragment header) is obtained by averaging the
* GstAdaptiveDemuxStream->last_download_time and is stored in
* GstDashDemuxStream->average_download_time. Those values include the
* network latency and full download time, which are more interesting and
* correct than just bitrates (with small download sizes, the impact of the
* network latency is much higher).
*
* The current position is calculated based on the fragment timestamp and
* the current keyframe index within that fragment. It is stored in
* GstDashDemuxStream->actual_position.
*
* The downstream position of the pipeline is obtained via QoS events and
* is stored in GstAdaptiveDemux (note: it's a running time value).
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
*
* The estimated buffering level between dashdemux and downstream is
* therefore:
* buffering_level = running_time(actual_position) - qos_earliest_time
*
* In order to avoid rebuffering, we want to ensure that the next keyframe
* (including potential fragment header) we request will be download, demuxed
* and decoded in time so that it is not late. That next keyframe time is
* called the "target_time" and is calculated whenever we have finished
* pushing a keyframe downstream.
*
* One simple observation at this point is that we *need* to make sure that
* the target time is chosen such that:
* running_time(target_time) > qos_earliest_time + average_download_time
*
* i.e. we chose a target time which will be greater than the time at which
* downstream will be once we request and download the keyframe (otherwise
* we're guaranteed to be late).
*
* This would provide the highest number of displayed frames per
* second, but it is just a *minimal* value and is not enough as-is,
* since it doesn't take into account the following items which could
* cause frames to arrive late (and therefore rebuffering):
* * Network jitter (i.e. by how much the download time can fluctuate)
* * Network stalling
* * Different keyframe sizes (and therefore download time)
* * Decoding speed
*
* Instead, we adjust the target time calculation based on the
* buffering_level.
*
* The smaller the buffering level is (i.e. the closer we are between
2019-09-02 19:08:44 +00:00
* current and downstream), the more aggressively we skip forward (and
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
* guarantee the keyframe will be downloaded, decoded and displayed in
* time). And the higher the buffering level, the least aggresivelly
* we need to skip forward (and therefore display more frames per
* second).
*
2019-09-02 19:08:44 +00:00
* Right now the threshold for aggressive switching is set to 3
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
* average_download_time. Below that buffering level we set the target time
* to at least 3 average_download_time distance beyond the
* qos_earliest_time.
*
* If we are above that buffering level we set the target time to:
* position(running_time(position) + average_download_time)
*
* The logic is therefore:
* WHILE(!EOS)
* Calculate target_time
* Advance to keyframe/fragment for that target_time
* Adaptivedemux downloads that keyframe/fragment
*
*/
2013-05-08 14:13:32 +00:00
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif
#include <string.h>
#include <stdio.h>
#include <stdlib.h>
#include <inttypes.h>
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
#include <gio/gio.h>
2013-05-08 14:13:32 +00:00
#include <gst/base/gsttypefindhelper.h>
2014-08-26 19:45:46 +00:00
#include <gst/tag/tag.h>
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
#include <gst/net/gstnet.h>
#include <glib/gi18n-lib.h>
2013-05-08 14:13:32 +00:00
#include "gstdashdemux.h"
#include "gstdash_debug.h"
2013-05-08 14:13:32 +00:00
static GstStaticPadTemplate gst_dash_demux_videosrc_template =
GST_STATIC_PAD_TEMPLATE ("video_%02u",
GST_PAD_SRC,
GST_PAD_SOMETIMES,
GST_STATIC_CAPS_ANY);
static GstStaticPadTemplate gst_dash_demux_audiosrc_template =
GST_STATIC_PAD_TEMPLATE ("audio_%02u",
2013-05-08 14:13:32 +00:00
GST_PAD_SRC,
GST_PAD_SOMETIMES,
GST_STATIC_CAPS_ANY);
static GstStaticPadTemplate gst_dash_demux_subtitlesrc_template =
GST_STATIC_PAD_TEMPLATE ("subtitle_%02u",
GST_PAD_SRC,
GST_PAD_SOMETIMES,
GST_STATIC_CAPS_ANY);
2013-05-08 14:13:32 +00:00
static GstStaticPadTemplate sinktemplate = GST_STATIC_PAD_TEMPLATE ("sink",
GST_PAD_SINK,
GST_PAD_ALWAYS,
GST_STATIC_CAPS ("application/dash+xml"));
2013-05-08 14:13:32 +00:00
GST_DEBUG_CATEGORY (gst_dash_demux_debug);
2013-05-08 14:13:32 +00:00
#define GST_CAT_DEFAULT gst_dash_demux_debug
enum
{
PROP_0,
PROP_MAX_BUFFERING_TIME,
PROP_BANDWIDTH_USAGE,
PROP_MAX_BITRATE,
PROP_MAX_VIDEO_WIDTH,
PROP_MAX_VIDEO_HEIGHT,
PROP_MAX_VIDEO_FRAMERATE,
PROP_PRESENTATION_DELAY,
2013-05-08 14:13:32 +00:00
PROP_LAST
};
/* Default values for properties */
2012-11-21 13:14:34 +00:00
#define DEFAULT_MAX_BUFFERING_TIME 30 /* in seconds */
#define DEFAULT_BANDWIDTH_USAGE 0.8f /* 0 to 1 */
#define DEFAULT_MAX_BITRATE 0 /* in bit/s */
#define DEFAULT_MAX_VIDEO_WIDTH 0
#define DEFAULT_MAX_VIDEO_HEIGHT 0
#define DEFAULT_MAX_VIDEO_FRAMERATE_N 0
#define DEFAULT_MAX_VIDEO_FRAMERATE_D 1
#define DEFAULT_PRESENTATION_DELAY "10s" /* 10s */
2013-05-08 14:13:32 +00:00
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
/* Clock drift compensation for live streams */
#define SLOW_CLOCK_UPDATE_INTERVAL (1000000 * 30 * 60) /* 30 minutes */
#define FAST_CLOCK_UPDATE_INTERVAL (1000000 * 30) /* 30 seconds */
#define SUPPORTED_CLOCK_FORMATS (GST_MPD_UTCTIMING_TYPE_NTP | GST_MPD_UTCTIMING_TYPE_HTTP_HEAD | GST_MPD_UTCTIMING_TYPE_HTTP_XSDATE | GST_MPD_UTCTIMING_TYPE_HTTP_ISO | GST_MPD_UTCTIMING_TYPE_HTTP_NTP)
#define NTP_TO_UNIX_EPOCH G_GUINT64_CONSTANT(2208988800) /* difference (in seconds) between NTP epoch and Unix epoch */
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
struct _GstDashDemuxClockDrift
{
GMutex clock_lock; /* used to protect access to struct */
guint selected_url;
gint64 next_update;
/* @clock_compensation: amount (in usecs) to add to client's idea of
now to map it to the server's idea of now */
GTimeSpan clock_compensation;
GstClock *ntp_clock;
};
typedef struct
{
guint64 start_offset, end_offset;
/* TODO: Timestamp and duration */
} GstDashStreamSyncSample;
2013-05-08 14:13:32 +00:00
/* GObject */
static void gst_dash_demux_set_property (GObject * object, guint prop_id,
const GValue * value, GParamSpec * pspec);
static void gst_dash_demux_get_property (GObject * object, guint prop_id,
GValue * value, GParamSpec * pspec);
static void gst_dash_demux_dispose (GObject * obj);
2014-08-26 19:45:46 +00:00
/* GstAdaptiveDemux */
static GstClockTime gst_dash_demux_get_duration (GstAdaptiveDemux * ademux);
static gboolean gst_dash_demux_is_live (GstAdaptiveDemux * ademux);
static void gst_dash_demux_reset (GstAdaptiveDemux * ademux);
static gboolean gst_dash_demux_process_manifest (GstAdaptiveDemux * ademux,
GstBuffer * buf);
static gboolean gst_dash_demux_seek (GstAdaptiveDemux * demux, GstEvent * seek);
static GstFlowReturn
gst_dash_demux_stream_update_fragment_info (GstAdaptiveDemuxStream * stream);
static GstFlowReturn gst_dash_demux_stream_seek (GstAdaptiveDemuxStream *
stream, gboolean forward, GstSeekFlags flags, GstClockTime ts,
GstClockTime * final_ts);
static gboolean gst_dash_demux_stream_has_next_fragment (GstAdaptiveDemuxStream
* stream);
2014-08-26 19:45:46 +00:00
static GstFlowReturn
gst_dash_demux_stream_advance_fragment (GstAdaptiveDemuxStream * stream);
static gboolean
gst_dash_demux_stream_advance_subfragment (GstAdaptiveDemuxStream * stream);
2014-08-26 19:45:46 +00:00
static gboolean gst_dash_demux_stream_select_bitrate (GstAdaptiveDemuxStream *
stream, guint64 bitrate);
static gint64 gst_dash_demux_get_manifest_update_interval (GstAdaptiveDemux *
demux);
static GstFlowReturn gst_dash_demux_update_manifest_data (GstAdaptiveDemux *
demux, GstBuffer * buf);
2014-08-26 19:45:46 +00:00
static gint64
gst_dash_demux_stream_get_fragment_waiting_time (GstAdaptiveDemuxStream *
stream);
static void gst_dash_demux_advance_period (GstAdaptiveDemux * demux);
static gboolean gst_dash_demux_has_next_period (GstAdaptiveDemux * demux);
static GstFlowReturn gst_dash_demux_data_received (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream, GstBuffer * buffer);
static gboolean
gst_dash_demux_stream_fragment_start (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream);
static GstFlowReturn
gst_dash_demux_stream_fragment_finished (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream);
static gboolean gst_dash_demux_need_another_chunk (GstAdaptiveDemuxStream *
stream);
2013-05-08 14:13:32 +00:00
/* GstDashDemux */
2012-12-20 08:04:28 +00:00
static gboolean gst_dash_demux_setup_all_streams (GstDashDemux * demux);
static void gst_dash_demux_stream_free (GstAdaptiveDemuxStream * stream);
2014-08-26 19:45:46 +00:00
static GstCaps *gst_dash_demux_get_input_caps (GstDashDemux * demux,
GstActiveStream * stream);
static GstPad *gst_dash_demux_create_pad (GstDashDemux * demux,
GstActiveStream * stream);
static GstDashDemuxClockDrift *gst_dash_demux_clock_drift_new (GstDashDemux *
demux);
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
static void gst_dash_demux_clock_drift_free (GstDashDemuxClockDrift *);
static gboolean gst_dash_demux_poll_clock_drift (GstDashDemux * demux);
static GTimeSpan gst_dash_demux_get_clock_compensation (GstDashDemux * demux);
static GDateTime *gst_dash_demux_get_server_now_utc (GstDashDemux * demux);
2013-05-08 14:13:32 +00:00
#define SIDX(s) (&(s)->sidx_parser.sidx)
static inline GstSidxBoxEntry *
SIDX_ENTRY (GstDashDemuxStream * s, gint i)
{
g_assert (i < SIDX (s)->entries_count);
return &(SIDX (s)->entries[(i)]);
}
#define SIDX_CURRENT_ENTRY(s) SIDX_ENTRY(s, SIDX(s)->entry_index)
static void gst_dash_demux_send_content_protection_event (gpointer cp_data,
gpointer stream);
2013-04-18 10:41:09 +00:00
#define gst_dash_demux_parent_class parent_class
2014-08-26 19:45:46 +00:00
G_DEFINE_TYPE_WITH_CODE (GstDashDemux, gst_dash_demux, GST_TYPE_ADAPTIVE_DEMUX,
2013-04-18 10:41:09 +00:00
GST_DEBUG_CATEGORY_INIT (gst_dash_demux_debug, "dashdemux", 0,
"dashdemux element")
2014-08-26 19:45:46 +00:00
);
GST_ELEMENT_REGISTER_DEFINE (dashdemux, "dashdemux", GST_RANK_PRIMARY,
GST_TYPE_DASH_DEMUX);
2013-05-08 14:13:32 +00:00
static void
gst_dash_demux_dispose (GObject * obj)
{
GstDashDemux *demux = GST_DASH_DEMUX (obj);
2014-08-26 19:45:46 +00:00
gst_dash_demux_reset (GST_ADAPTIVE_DEMUX_CAST (demux));
2014-08-26 19:45:46 +00:00
if (demux->client) {
gst_mpd_client_free (demux->client);
demux->client = NULL;
2013-05-08 14:13:32 +00:00
}
g_mutex_clear (&demux->client_lock);
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
gst_dash_demux_clock_drift_free (demux->clock_drift);
demux->clock_drift = NULL;
g_free (demux->default_presentation_delay);
2013-05-08 14:13:32 +00:00
G_OBJECT_CLASS (parent_class)->dispose (obj);
}
static gboolean
gst_dash_demux_get_live_seek_range (GstAdaptiveDemux * demux, gint64 * start,
gint64 * stop)
{
GstDashDemux *self = GST_DASH_DEMUX (demux);
GDateTime *now;
GDateTime *mstart;
GTimeSpan stream_now;
GstClockTime seg_duration;
if (self->client->mpd_root_node->availabilityStartTime == NULL)
return FALSE;
seg_duration = gst_mpd_client_get_maximum_segment_duration (self->client);
now = gst_dash_demux_get_server_now_utc (self);
mstart =
gst_date_time_to_g_date_time (self->client->mpd_root_node->
availabilityStartTime);
stream_now = g_date_time_difference (now, mstart);
g_date_time_unref (now);
g_date_time_unref (mstart);
if (stream_now <= 0)
return FALSE;
*stop = stream_now * GST_USECOND;
if (self->client->mpd_root_node->timeShiftBufferDepth ==
GST_MPD_DURATION_NONE) {
*start = 0;
} else {
*start =
*stop -
(self->client->mpd_root_node->timeShiftBufferDepth * GST_MSECOND);
if (*start < 0)
*start = 0;
}
/* As defined in 5.3.9.5.3 of the DASH specification, a segment does
not become available until the sum of:
* the value of the MPD@availabilityStartTime,
* the PeriodStart time of the containing Period
* the MPD start time of the Media Segment, and
* the MPD duration of the Media Segment.
Therefore we need to subtract the media segment duration from the stop
time.
*/
*stop -= seg_duration;
return TRUE;
}
static GstClockTime
gst_dash_demux_get_presentation_offset (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream)
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
return gst_mpd_client_get_stream_presentation_offset (dashdemux->client,
dashstream->index);
}
static GstClockTime
gst_dash_demux_get_period_start_time (GstAdaptiveDemux * demux)
{
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
return gst_mpd_client_get_period_start_time (dashdemux->client);
}
2013-05-08 14:13:32 +00:00
static void
gst_dash_demux_class_init (GstDashDemuxClass * klass)
{
GObjectClass *gobject_class;
GstElementClass *gstelement_class;
2014-08-26 19:45:46 +00:00
GstAdaptiveDemuxClass *gstadaptivedemux_class;
2013-05-08 14:13:32 +00:00
gobject_class = (GObjectClass *) klass;
gstelement_class = (GstElementClass *) klass;
2014-08-26 19:45:46 +00:00
gstadaptivedemux_class = (GstAdaptiveDemuxClass *) klass;
2013-05-08 14:13:32 +00:00
gobject_class->set_property = gst_dash_demux_set_property;
gobject_class->get_property = gst_dash_demux_get_property;
gobject_class->dispose = gst_dash_demux_dispose;
#ifndef GST_REMOVE_DEPRECATED
2013-05-08 14:13:32 +00:00
g_object_class_install_property (gobject_class, PROP_MAX_BUFFERING_TIME,
g_param_spec_uint ("max-buffering-time", "Maximum buffering time",
"Maximum number of seconds of buffer accumulated during playback"
"(deprecated)",
2013-05-08 14:13:32 +00:00
2, G_MAXUINT, DEFAULT_MAX_BUFFERING_TIME,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS | G_PARAM_DEPRECATED));
2013-05-08 14:13:32 +00:00
g_object_class_install_property (gobject_class, PROP_BANDWIDTH_USAGE,
g_param_spec_float ("bandwidth-usage",
"Bandwidth usage [0..1]",
"Percentage of the available bandwidth to use when "
"selecting representations (deprecated)",
2013-05-08 14:13:32 +00:00
0, 1, DEFAULT_BANDWIDTH_USAGE,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS | G_PARAM_DEPRECATED));
#endif
2013-05-08 14:13:32 +00:00
g_object_class_install_property (gobject_class, PROP_MAX_BITRATE,
g_param_spec_uint ("max-bitrate", "Max bitrate",
"Max of bitrate supported by target video decoder (0 = no maximum)",
0, G_MAXUINT, DEFAULT_MAX_BITRATE,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
g_object_class_install_property (gobject_class, PROP_MAX_VIDEO_WIDTH,
g_param_spec_uint ("max-video-width", "Max video width",
"Max video width to select (0 = no maximum)",
0, G_MAXUINT, DEFAULT_MAX_VIDEO_WIDTH,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
g_object_class_install_property (gobject_class, PROP_MAX_VIDEO_HEIGHT,
g_param_spec_uint ("max-video-height", "Max video height",
"Max video height to select (0 = no maximum)",
0, G_MAXUINT, DEFAULT_MAX_VIDEO_HEIGHT,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
g_object_class_install_property (gobject_class, PROP_MAX_VIDEO_FRAMERATE,
gst_param_spec_fraction ("max-video-framerate", "Max video framerate",
"Max video framerate to select (0/1 = no maximum)",
0, 1, G_MAXINT, 1, DEFAULT_MAX_VIDEO_FRAMERATE_N,
DEFAULT_MAX_VIDEO_FRAMERATE_D,
2013-05-08 14:13:32 +00:00
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
g_object_class_install_property (gobject_class, PROP_PRESENTATION_DELAY,
g_param_spec_string ("presentation-delay", "Presentation delay",
"Default presentation delay (in seconds, milliseconds or fragments) (e.g. 12s, 2500ms, 3f)",
DEFAULT_PRESENTATION_DELAY,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
gst_element_class_add_static_pad_template (gstelement_class,
&gst_dash_demux_audiosrc_template);
gst_element_class_add_static_pad_template (gstelement_class,
&gst_dash_demux_videosrc_template);
gst_element_class_add_static_pad_template (gstelement_class,
&gst_dash_demux_subtitlesrc_template);
gst_element_class_add_static_pad_template (gstelement_class, &sinktemplate);
2013-04-18 10:41:09 +00:00
gst_element_class_set_static_metadata (gstelement_class,
"DASH Demuxer",
"Codec/Demuxer/Adaptive",
2013-04-18 10:41:09 +00:00
"Dynamic Adaptive Streaming over HTTP demuxer",
"David Corvoysier <david.corvoysier@orange.com>, "
"Hamid Zakari <hamid.zakari@gmail.com>, "
"Gianluca Gennari <gennarone@gmail.com>");
2014-08-26 19:45:46 +00:00
gstadaptivedemux_class->get_duration = gst_dash_demux_get_duration;
gstadaptivedemux_class->is_live = gst_dash_demux_is_live;
gstadaptivedemux_class->reset = gst_dash_demux_reset;
gstadaptivedemux_class->seek = gst_dash_demux_seek;
gstadaptivedemux_class->process_manifest = gst_dash_demux_process_manifest;
gstadaptivedemux_class->update_manifest_data =
gst_dash_demux_update_manifest_data;
2014-08-26 19:45:46 +00:00
gstadaptivedemux_class->get_manifest_update_interval =
gst_dash_demux_get_manifest_update_interval;
gstadaptivedemux_class->has_next_period = gst_dash_demux_has_next_period;
gstadaptivedemux_class->advance_period = gst_dash_demux_advance_period;
gstadaptivedemux_class->stream_has_next_fragment =
gst_dash_demux_stream_has_next_fragment;
2014-08-26 19:45:46 +00:00
gstadaptivedemux_class->stream_advance_fragment =
gst_dash_demux_stream_advance_fragment;
gstadaptivedemux_class->stream_get_fragment_waiting_time =
gst_dash_demux_stream_get_fragment_waiting_time;
gstadaptivedemux_class->stream_seek = gst_dash_demux_stream_seek;
gstadaptivedemux_class->stream_select_bitrate =
gst_dash_demux_stream_select_bitrate;
gstadaptivedemux_class->stream_update_fragment_info =
gst_dash_demux_stream_update_fragment_info;
gstadaptivedemux_class->stream_free = gst_dash_demux_stream_free;
gstadaptivedemux_class->get_live_seek_range =
gst_dash_demux_get_live_seek_range;
gstadaptivedemux_class->get_presentation_offset =
gst_dash_demux_get_presentation_offset;
gstadaptivedemux_class->get_period_start_time =
gst_dash_demux_get_period_start_time;
gstadaptivedemux_class->start_fragment = gst_dash_demux_stream_fragment_start;
gstadaptivedemux_class->finish_fragment =
gst_dash_demux_stream_fragment_finished;
gstadaptivedemux_class->data_received = gst_dash_demux_data_received;
gstadaptivedemux_class->need_another_chunk =
gst_dash_demux_need_another_chunk;
2013-05-08 14:13:32 +00:00
}
static void
2013-04-18 10:41:09 +00:00
gst_dash_demux_init (GstDashDemux * demux)
2013-05-08 14:13:32 +00:00
{
/* Properties */
demux->max_buffering_time = DEFAULT_MAX_BUFFERING_TIME * GST_SECOND;
demux->max_bitrate = DEFAULT_MAX_BITRATE;
demux->max_video_width = DEFAULT_MAX_VIDEO_WIDTH;
demux->max_video_height = DEFAULT_MAX_VIDEO_HEIGHT;
demux->max_video_framerate_n = DEFAULT_MAX_VIDEO_FRAMERATE_N;
demux->max_video_framerate_d = DEFAULT_MAX_VIDEO_FRAMERATE_D;
demux->default_presentation_delay = g_strdup (DEFAULT_PRESENTATION_DELAY);
2013-05-08 14:13:32 +00:00
g_mutex_init (&demux->client_lock);
2014-08-26 19:45:46 +00:00
gst_adaptive_demux_set_stream_struct_size (GST_ADAPTIVE_DEMUX_CAST (demux),
sizeof (GstDashDemuxStream));
2013-05-08 14:13:32 +00:00
}
static void
gst_dash_demux_set_property (GObject * object, guint prop_id,
const GValue * value, GParamSpec * pspec)
{
GstAdaptiveDemux *adaptivedemux = GST_ADAPTIVE_DEMUX_CAST (object);
2013-05-08 14:13:32 +00:00
GstDashDemux *demux = GST_DASH_DEMUX (object);
switch (prop_id) {
case PROP_MAX_BUFFERING_TIME:
demux->max_buffering_time = g_value_get_uint (value) * GST_SECOND;
break;
case PROP_BANDWIDTH_USAGE:
adaptivedemux->bitrate_limit = g_value_get_float (value);
2013-05-08 14:13:32 +00:00
break;
case PROP_MAX_BITRATE:
demux->max_bitrate = g_value_get_uint (value);
break;
case PROP_MAX_VIDEO_WIDTH:
demux->max_video_width = g_value_get_uint (value);
break;
case PROP_MAX_VIDEO_HEIGHT:
demux->max_video_height = g_value_get_uint (value);
break;
case PROP_MAX_VIDEO_FRAMERATE:
demux->max_video_framerate_n = gst_value_get_fraction_numerator (value);
demux->max_video_framerate_d = gst_value_get_fraction_denominator (value);
break;
case PROP_PRESENTATION_DELAY:
g_free (demux->default_presentation_delay);
demux->default_presentation_delay = g_value_dup_string (value);
break;
2013-05-08 14:13:32 +00:00
default:
G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
break;
}
}
static void
gst_dash_demux_get_property (GObject * object, guint prop_id, GValue * value,
GParamSpec * pspec)
{
GstAdaptiveDemux *adaptivedemux = GST_ADAPTIVE_DEMUX_CAST (object);
2013-05-08 14:13:32 +00:00
GstDashDemux *demux = GST_DASH_DEMUX (object);
switch (prop_id) {
case PROP_MAX_BUFFERING_TIME:
g_value_set_uint (value, demux->max_buffering_time / GST_SECOND);
2013-05-08 14:13:32 +00:00
break;
case PROP_BANDWIDTH_USAGE:
g_value_set_float (value, adaptivedemux->bitrate_limit);
2013-05-08 14:13:32 +00:00
break;
case PROP_MAX_BITRATE:
g_value_set_uint (value, demux->max_bitrate);
break;
case PROP_MAX_VIDEO_WIDTH:
g_value_set_uint (value, demux->max_video_width);
break;
case PROP_MAX_VIDEO_HEIGHT:
g_value_set_uint (value, demux->max_video_height);
break;
case PROP_MAX_VIDEO_FRAMERATE:
gst_value_set_fraction (value, demux->max_video_framerate_n,
demux->max_video_framerate_d);
break;
case PROP_PRESENTATION_DELAY:
if (demux->default_presentation_delay == NULL)
g_value_set_static_string (value, "");
else
g_value_set_string (value, demux->default_presentation_delay);
break;
2013-05-08 14:13:32 +00:00
default:
G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
break;
}
}
static gboolean
gst_dash_demux_setup_mpdparser_streams (GstDashDemux * demux,
GstMPDClient * client)
{
gboolean has_streams = FALSE;
GList *adapt_sets, *iter;
adapt_sets = gst_mpd_client_get_adaptation_sets (client);
for (iter = adapt_sets; iter; iter = g_list_next (iter)) {
GstMPDAdaptationSetNode *adapt_set_node = iter->data;
gst_mpd_client_setup_streaming (client, adapt_set_node);
has_streams = TRUE;
}
if (!has_streams) {
GST_ELEMENT_ERROR (demux, STREAM, DEMUX, ("Manifest has no playable "
"streams"), ("No streams could be activated from the manifest"));
}
return has_streams;
}
static gboolean
gst_dash_demux_setup_all_streams (GstDashDemux * demux)
{
guint i;
GST_DEBUG_OBJECT (demux, "Setting up streams for period %d",
gst_mpd_client_get_period_index (demux->client));
/* clean old active stream list, if any */
gst_mpd_client_active_streams_free (demux->client);
if (!gst_dash_demux_setup_mpdparser_streams (demux, demux->client)) {
return FALSE;
}
GST_DEBUG_OBJECT (demux, "Creating stream objects");
for (i = 0; i < gst_mpd_client_get_nb_active_stream (demux->client); i++) {
GstDashDemuxStream *stream;
GstActiveStream *active_stream;
GstCaps *caps;
GstStructure *s;
2014-08-26 19:45:46 +00:00
GstPad *srcpad;
gchar *lang = NULL;
2014-08-26 19:45:46 +00:00
GstTagList *tags = NULL;
active_stream =
gst_mpd_client_get_active_stream_by_index (demux->client, i);
if (active_stream == NULL)
continue;
if (demux->trickmode_no_audio
&& active_stream->mimeType == GST_STREAM_AUDIO) {
GST_DEBUG_OBJECT (demux,
"Skipping audio stream %d because of TRICKMODE_NO_AUDIO flag", i);
continue;
}
srcpad = gst_dash_demux_create_pad (demux, active_stream);
if (srcpad == NULL)
continue;
caps = gst_dash_demux_get_input_caps (demux, active_stream);
GST_LOG_OBJECT (demux, "Creating stream %d %" GST_PTR_FORMAT, i, caps);
if (active_stream->cur_adapt_set) {
GstMPDAdaptationSetNode *adp_set = active_stream->cur_adapt_set;
lang = adp_set->lang;
/* Fallback to the language in ContentComponent node */
if (lang == NULL) {
GList *it;
for (it = adp_set->ContentComponents; it; it = it->next) {
GstMPDContentComponentNode *cc_node = it->data;
if (cc_node->lang) {
lang = cc_node->lang;
break;
}
}
}
}
if (lang) {
if (gst_tag_check_language_code (lang))
tags = gst_tag_list_new (GST_TAG_LANGUAGE_CODE, lang, NULL);
else
tags = gst_tag_list_new (GST_TAG_LANGUAGE_NAME, lang, NULL);
}
2014-08-26 19:45:46 +00:00
stream = (GstDashDemuxStream *)
gst_adaptive_demux_stream_new (GST_ADAPTIVE_DEMUX_CAST (demux), srcpad);
stream->active_stream = active_stream;
if (active_stream->cur_representation) {
stream->last_representation_id =
g_strdup (stream->active_stream->cur_representation->id);
} else {
stream->last_representation_id = NULL;
}
s = gst_caps_get_structure (caps, 0);
stream->allow_sidx =
gst_mpd_client_has_isoff_ondemand_profile (demux->client);
stream->is_isobmff = gst_structure_has_name (s, "video/quicktime")
|| gst_structure_has_name (s, "audio/x-m4a");
stream->first_sync_sample_always_after_moof = TRUE;
stream->adapter = gst_adapter_new ();
2014-08-26 19:45:46 +00:00
gst_adaptive_demux_stream_set_caps (GST_ADAPTIVE_DEMUX_STREAM_CAST (stream),
caps);
if (tags)
gst_adaptive_demux_stream_set_tags (GST_ADAPTIVE_DEMUX_STREAM_CAST
(stream), tags);
stream->index = i;
stream->pending_seek_ts = GST_CLOCK_TIME_NONE;
stream->sidx_position = GST_CLOCK_TIME_NONE;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
stream->actual_position = GST_CLOCK_TIME_NONE;
stream->target_time = GST_CLOCK_TIME_NONE;
/* Set a default average keyframe download time of a quarter of a second */
stream->average_download_time = 250 * GST_MSECOND;
if (active_stream->cur_adapt_set &&
GST_MPD_REPRESENTATION_BASE_NODE (active_stream->
cur_adapt_set)->ContentProtection) {
GST_DEBUG_OBJECT (demux, "Adding ContentProtection events to source pad");
g_list_foreach (GST_MPD_REPRESENTATION_BASE_NODE
(active_stream->cur_adapt_set)->ContentProtection,
gst_dash_demux_send_content_protection_event, stream);
}
gst_isoff_sidx_parser_init (&stream->sidx_parser);
2014-08-26 19:45:46 +00:00
}
return TRUE;
}
static void
gst_dash_demux_send_content_protection_event (gpointer data, gpointer userdata)
{
GstMPDDescriptorTypeNode *cp = (GstMPDDescriptorTypeNode *) data;
GstDashDemuxStream *stream = (GstDashDemuxStream *) userdata;
GstEvent *event;
GstBuffer *pssi;
glong pssi_len;
gchar *schemeIdUri;
GstPad *pad = GST_ADAPTIVE_DEMUX_STREAM_PAD (stream);
if (cp->schemeIdUri == NULL)
return;
GST_TRACE_OBJECT (pad, "check schemeIdUri %s", cp->schemeIdUri);
/* RFC 2141 states: The leading "urn:" sequence is case-insensitive */
schemeIdUri = g_ascii_strdown (cp->schemeIdUri, -1);
if (g_str_has_prefix (schemeIdUri, "urn:uuid:")) {
pssi_len = strlen (cp->value);
pssi = gst_buffer_new_memdup (cp->value, pssi_len);
/* RFC 4122 states that the hex part of a UUID is in lower case,
* but some streams seem to ignore this and use upper case for the
* protection system ID */
event = gst_event_new_protection (cp->schemeIdUri + 9, pssi, "dash/mpd");
GST_LOG_OBJECT (pad,
"Queueing protection event %" GST_PTR_FORMAT " on source pad", event);
gst_adaptive_demux_stream_queue_event ((GstAdaptiveDemuxStream *) stream,
event);
gst_buffer_unref (pssi);
}
g_free (schemeIdUri);
}
2014-08-26 19:45:46 +00:00
static GstClockTime
gst_dash_demux_get_duration (GstAdaptiveDemux * ademux)
2013-05-08 14:13:32 +00:00
{
2014-08-26 19:45:46 +00:00
GstDashDemux *demux = GST_DASH_DEMUX_CAST (ademux);
2014-08-26 19:45:46 +00:00
g_return_val_if_fail (demux->client != NULL, GST_CLOCK_TIME_NONE);
2014-08-26 19:45:46 +00:00
return gst_mpd_client_get_media_presentation_duration (demux->client);
}
2014-08-26 19:45:46 +00:00
static gboolean
gst_dash_demux_is_live (GstAdaptiveDemux * ademux)
{
GstDashDemux *demux = GST_DASH_DEMUX_CAST (ademux);
2014-08-26 19:45:46 +00:00
g_return_val_if_fail (demux->client != NULL, FALSE);
2013-05-08 14:13:32 +00:00
2014-08-26 19:45:46 +00:00
return gst_mpd_client_is_live (demux->client);
2013-05-08 14:13:32 +00:00
}
static gboolean
2014-08-26 19:45:46 +00:00
gst_dash_demux_setup_streams (GstAdaptiveDemux * demux)
2013-05-08 14:13:32 +00:00
{
2014-08-26 19:45:46 +00:00
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
gboolean ret = TRUE;
GstDateTime *now = NULL;
guint period_idx;
2014-08-26 19:45:46 +00:00
/* setup video, audio and subtitle streams, starting from first Period if
* non-live */
period_idx = 0;
if (gst_mpd_client_is_live (dashdemux->client)) {
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
GDateTime *g_now;
if (dashdemux->client->mpd_root_node->availabilityStartTime == NULL) {
ret = FALSE;
GST_ERROR_OBJECT (demux, "MPD does not have availabilityStartTime");
goto done;
}
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
if (dashdemux->clock_drift == NULL) {
gchar **urls;
urls =
gst_mpd_client_get_utc_timing_sources (dashdemux->client,
SUPPORTED_CLOCK_FORMATS, NULL);
if (urls) {
GST_DEBUG_OBJECT (dashdemux, "Found a supported UTCTiming element");
dashdemux->clock_drift = gst_dash_demux_clock_drift_new (dashdemux);
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
gst_dash_demux_poll_clock_drift (dashdemux);
}
}
/* get period index for period encompassing the current time */
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
g_now = gst_dash_demux_get_server_now_utc (dashdemux);
now = gst_date_time_new_from_g_date_time (g_now);
if (dashdemux->client->mpd_root_node->suggestedPresentationDelay != -1) {
GstDateTime *target = gst_mpd_client_add_time_difference (now,
dashdemux->client->mpd_root_node->suggestedPresentationDelay * -1000);
gst_date_time_unref (now);
now = target;
} else if (dashdemux->default_presentation_delay) {
gint64 dfp =
gst_mpd_client_parse_default_presentation_delay (dashdemux->client,
dashdemux->default_presentation_delay);
GstDateTime *target = gst_mpd_client_add_time_difference (now,
dfp * -1000);
gst_date_time_unref (now);
now = target;
}
period_idx =
gst_mpd_client_get_period_index_at_time (dashdemux->client, now);
if (period_idx == G_MAXUINT) {
#ifndef GST_DISABLE_GST_DEBUG
gchar *date_str = gst_date_time_to_iso8601_string (now);
GST_DEBUG_OBJECT (demux, "Unable to find live period active at %s",
date_str);
g_free (date_str);
#endif
ret = FALSE;
goto done;
}
}
if (!gst_mpd_client_set_period_index (dashdemux->client, period_idx) ||
2014-08-26 19:45:46 +00:00
!gst_dash_demux_setup_all_streams (dashdemux)) {
ret = FALSE;
goto done;
}
/* If stream is live, try to find the segment that
* is closest to current time */
if (gst_mpd_client_is_live (dashdemux->client)) {
GDateTime *gnow;
2014-08-26 19:45:46 +00:00
GST_DEBUG_OBJECT (demux, "Seeking to current time of day for live stream ");
gnow = gst_date_time_to_g_date_time (now);
gst_mpd_client_seek_to_time (dashdemux->client, gnow);
g_date_time_unref (gnow);
2014-08-26 19:45:46 +00:00
} else {
GST_DEBUG_OBJECT (demux, "Seeking to first segment for on-demand stream ");
2014-08-26 19:45:46 +00:00
/* start playing from the first segment */
gst_mpd_client_seek_to_first_segment (dashdemux->client);
2013-05-08 14:13:32 +00:00
}
2014-08-26 19:45:46 +00:00
done:
if (now != NULL)
gst_date_time_unref (now);
2013-05-08 14:13:32 +00:00
return ret;
}
2014-08-26 19:45:46 +00:00
static gboolean
gst_dash_demux_process_manifest (GstAdaptiveDemux * demux, GstBuffer * buf)
2013-05-08 14:13:32 +00:00
{
2014-08-26 19:45:46 +00:00
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
gboolean ret = FALSE;
gchar *manifest;
GstMapInfo mapinfo;
2013-05-08 14:13:32 +00:00
2014-08-26 19:45:46 +00:00
if (dashdemux->client)
gst_mpd_client_free (dashdemux->client);
dashdemux->client = gst_mpd_client_new ();
gst_mpd_client_set_uri_downloader (dashdemux->client, demux->downloader);
2013-05-08 14:13:32 +00:00
2014-08-26 19:45:46 +00:00
dashdemux->client->mpd_uri = g_strdup (demux->manifest_uri);
dashdemux->client->mpd_base_uri = g_strdup (demux->manifest_base_uri);
2014-08-26 19:45:46 +00:00
GST_DEBUG_OBJECT (demux, "Fetched MPD file at URI: %s (base: %s)",
dashdemux->client->mpd_uri,
GST_STR_NULL (dashdemux->client->mpd_base_uri));
2014-08-26 19:45:46 +00:00
if (gst_buffer_map (buf, &mapinfo, GST_MAP_READ)) {
manifest = (gchar *) mapinfo.data;
if (gst_mpd_client_parse (dashdemux->client, manifest, mapinfo.size)) {
if (gst_mpd_client_setup_media_presentation (dashdemux->client, 0, 0,
NULL)) {
2014-08-26 19:45:46 +00:00
ret = TRUE;
} else {
GST_ELEMENT_ERROR (demux, STREAM, DECODE,
("Incompatible manifest file."), (NULL));
}
}
gst_buffer_unmap (buf, &mapinfo);
} else {
GST_WARNING_OBJECT (demux, "Failed to map manifest buffer");
}
2013-05-08 14:13:32 +00:00
2014-08-26 19:45:46 +00:00
if (ret)
ret = gst_dash_demux_setup_streams (demux);
return ret;
2013-05-08 14:13:32 +00:00
}
static GstPad *
gst_dash_demux_create_pad (GstDashDemux * demux, GstActiveStream * stream)
{
GstPad *pad;
GstPadTemplate *tmpl;
gchar *name;
switch (stream->mimeType) {
case GST_STREAM_AUDIO:
name = g_strdup_printf ("audio_%02u", demux->n_audio_streams++);
tmpl = gst_static_pad_template_get (&gst_dash_demux_audiosrc_template);
break;
case GST_STREAM_VIDEO:
name = g_strdup_printf ("video_%02u", demux->n_video_streams++);
tmpl = gst_static_pad_template_get (&gst_dash_demux_videosrc_template);
break;
case GST_STREAM_APPLICATION:
if (gst_mpd_client_active_stream_contains_subtitles (stream)) {
name = g_strdup_printf ("subtitle_%02u", demux->n_subtitle_streams++);
tmpl =
gst_static_pad_template_get (&gst_dash_demux_subtitlesrc_template);
} else {
return NULL;
}
break;
default:
g_assert_not_reached ();
return NULL;
}
/* Create and activate new pads */
pad = gst_pad_new_from_template (tmpl, name);
g_free (name);
gst_object_unref (tmpl);
gst_pad_set_active (pad, TRUE);
GST_INFO_OBJECT (demux, "Creating srcpad %s:%s", GST_DEBUG_PAD_NAME (pad));
return pad;
}
2013-05-08 14:13:32 +00:00
static void
2014-08-26 19:45:46 +00:00
gst_dash_demux_reset (GstAdaptiveDemux * ademux)
2013-05-08 14:13:32 +00:00
{
2014-08-26 19:45:46 +00:00
GstDashDemux *demux = GST_DASH_DEMUX_CAST (ademux);
2014-08-26 19:45:46 +00:00
GST_DEBUG_OBJECT (demux, "Resetting demux");
2014-08-26 19:45:46 +00:00
demux->end_of_period = FALSE;
demux->end_of_manifest = FALSE;
2014-08-26 19:45:46 +00:00
if (demux->client) {
gst_mpd_client_free (demux->client);
demux->client = NULL;
2013-05-08 14:13:32 +00:00
}
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
gst_dash_demux_clock_drift_free (demux->clock_drift);
demux->clock_drift = NULL;
2014-08-26 19:45:46 +00:00
demux->client = gst_mpd_client_new ();
gst_mpd_client_set_uri_downloader (demux->client, ademux->downloader);
demux->n_audio_streams = 0;
demux->n_video_streams = 0;
demux->n_subtitle_streams = 0;
demux->trickmode_no_audio = FALSE;
demux->allow_trickmode_key_units = TRUE;
}
2014-08-26 19:45:46 +00:00
static GstCaps *
gst_dash_demux_get_video_input_caps (GstDashDemux * demux,
GstActiveStream * stream)
{
2014-08-26 19:45:46 +00:00
guint width = 0, height = 0;
gint fps_num = 0, fps_den = 1;
gboolean have_fps = FALSE;
2014-08-26 19:45:46 +00:00
GstCaps *caps = NULL;
2014-08-26 19:45:46 +00:00
if (stream == NULL)
return NULL;
2019-09-02 19:08:44 +00:00
/* if bitstreamSwitching is true we don't need to switch pads on resolution change */
2014-08-26 19:45:46 +00:00
if (!gst_mpd_client_get_bitstream_switching_flag (stream)) {
width = gst_mpd_client_get_video_stream_width (stream);
height = gst_mpd_client_get_video_stream_height (stream);
have_fps =
gst_mpd_client_get_video_stream_framerate (stream, &fps_num, &fps_den);
}
caps = gst_mpd_client_get_stream_caps (stream);
if (caps == NULL)
2014-08-26 19:45:46 +00:00
return NULL;
2014-08-26 19:45:46 +00:00
if (width > 0 && height > 0) {
gst_caps_set_simple (caps, "width", G_TYPE_INT, width, "height",
G_TYPE_INT, height, NULL);
}
if (have_fps) {
gst_caps_set_simple (caps, "framerate", GST_TYPE_FRACTION, fps_num,
fps_den, NULL);
}
2014-08-26 19:45:46 +00:00
return caps;
}
2014-08-26 19:45:46 +00:00
static GstCaps *
gst_dash_demux_get_audio_input_caps (GstDashDemux * demux,
GstActiveStream * stream)
{
2014-08-26 19:45:46 +00:00
guint rate = 0, channels = 0;
GstCaps *caps = NULL;
2014-08-26 19:45:46 +00:00
if (stream == NULL)
return NULL;
2019-09-02 19:08:44 +00:00
/* if bitstreamSwitching is true we don't need to switch pads on rate/channels change */
2014-08-26 19:45:46 +00:00
if (!gst_mpd_client_get_bitstream_switching_flag (stream)) {
channels = gst_mpd_client_get_audio_stream_num_channels (stream);
rate = gst_mpd_client_get_audio_stream_rate (stream);
}
caps = gst_mpd_client_get_stream_caps (stream);
if (caps == NULL)
2014-08-26 19:45:46 +00:00
return NULL;
2014-08-26 19:45:46 +00:00
if (rate > 0) {
gst_caps_set_simple (caps, "rate", G_TYPE_INT, rate, NULL);
}
if (channels > 0) {
gst_caps_set_simple (caps, "channels", G_TYPE_INT, channels, NULL);
}
2014-08-26 19:45:46 +00:00
return caps;
}
2014-08-26 19:45:46 +00:00
static GstCaps *
gst_dash_demux_get_application_input_caps (GstDashDemux * demux,
GstActiveStream * stream)
{
GstCaps *caps = NULL;
2014-08-26 19:45:46 +00:00
if (stream == NULL)
return NULL;
2013-05-08 14:13:32 +00:00
caps = gst_mpd_client_get_stream_caps (stream);
if (caps == NULL)
2014-08-26 19:45:46 +00:00
return NULL;
2014-08-26 19:45:46 +00:00
return caps;
2013-05-08 14:13:32 +00:00
}
2014-08-26 19:45:46 +00:00
static GstCaps *
gst_dash_demux_get_input_caps (GstDashDemux * demux, GstActiveStream * stream)
{
2014-08-26 19:45:46 +00:00
switch (stream->mimeType) {
case GST_STREAM_VIDEO:
return gst_dash_demux_get_video_input_caps (demux, stream);
case GST_STREAM_AUDIO:
return gst_dash_demux_get_audio_input_caps (demux, stream);
case GST_STREAM_APPLICATION:
return gst_dash_demux_get_application_input_caps (demux, stream);
default:
return GST_CAPS_NONE;
}
}
static void
gst_dash_demux_stream_update_headers_info (GstAdaptiveDemuxStream * stream)
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (stream->demux);
gchar *path = NULL;
gst_mpd_client_get_next_header (dashdemux->client,
&path, dashstream->index,
&stream->fragment.header_range_start, &stream->fragment.header_range_end);
if (path != NULL) {
stream->fragment.header_uri =
gst_uri_join_strings (gst_mpd_client_get_baseURL (dashdemux->client,
dashstream->index), path);
g_free (path);
path = NULL;
}
gst_mpd_client_get_next_header_index (dashdemux->client,
&path, dashstream->index,
&stream->fragment.index_range_start, &stream->fragment.index_range_end);
if (path != NULL) {
stream->fragment.index_uri =
gst_uri_join_strings (gst_mpd_client_get_baseURL (dashdemux->client,
dashstream->index), path);
g_free (path);
}
}
static GstFlowReturn
2014-08-26 19:45:46 +00:00
gst_dash_demux_stream_update_fragment_info (GstAdaptiveDemuxStream * stream)
{
2014-08-26 19:45:46 +00:00
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (stream->demux);
GstClockTime ts;
GstMediaFragmentInfo fragment;
gboolean isombff;
2014-08-26 19:45:46 +00:00
gst_adaptive_demux_stream_fragment_clear (&stream->fragment);
isombff = gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client);
/* Reset chunk size if any */
stream->fragment.chunk_size = 0;
dashstream->current_fragment_keyframe_distance = GST_CLOCK_TIME_NONE;
if (GST_ADAPTIVE_DEMUX_STREAM_NEED_HEADER (stream) && isombff) {
gst_dash_demux_stream_update_headers_info (stream);
/* sidx entries may not be available in here */
if (stream->fragment.index_uri
&& dashstream->sidx_position != GST_CLOCK_TIME_NONE) {
/* request only the index to be downloaded as we need to reposition the
* stream to a subsegment */
return GST_FLOW_OK;
}
}
if (dashstream->moof_sync_samples
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (dashdemux)) {
GstDashStreamSyncSample *sync_sample =
&g_array_index (dashstream->moof_sync_samples, GstDashStreamSyncSample,
dashstream->current_sync_sample);
gst_mpd_client_get_next_fragment (dashdemux->client, dashstream->index,
&fragment);
if (isombff && dashstream->sidx_position != GST_CLOCK_TIME_NONE
&& SIDX (dashstream)->entries) {
GstSidxBoxEntry *entry = SIDX_CURRENT_ENTRY (dashstream);
dashstream->current_fragment_timestamp = fragment.timestamp = entry->pts;
dashstream->current_fragment_duration = fragment.duration =
entry->duration;
} else {
dashstream->current_fragment_timestamp = fragment.timestamp;
dashstream->current_fragment_duration = fragment.duration;
}
dashstream->current_fragment_keyframe_distance =
fragment.duration / dashstream->moof_sync_samples->len;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
dashstream->actual_position =
fragment.timestamp +
dashstream->current_sync_sample *
dashstream->current_fragment_keyframe_distance;
if (stream->segment.rate < 0.0)
dashstream->actual_position +=
dashstream->current_fragment_keyframe_distance;
dashstream->actual_position =
MIN (dashstream->actual_position,
fragment.timestamp + fragment.duration);
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
stream->fragment.uri = fragment.uri;
stream->fragment.timestamp = GST_CLOCK_TIME_NONE;
stream->fragment.duration = GST_CLOCK_TIME_NONE;
stream->fragment.range_start = sync_sample->start_offset;
stream->fragment.range_end = sync_sample->end_offset;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_DEBUG_OBJECT (stream->pad, "Actual position %" GST_TIME_FORMAT,
GST_TIME_ARGS (dashstream->actual_position));
return GST_FLOW_OK;
}
2014-08-26 19:45:46 +00:00
if (gst_mpd_client_get_next_fragment_timestamp (dashdemux->client,
dashstream->index, &ts)) {
/* For live streams, check whether the underlying representation changed
* (due to a manifest update with no matching representation) */
if (gst_mpd_client_is_live (dashdemux->client)
&& !GST_ADAPTIVE_DEMUX_STREAM_NEED_HEADER (stream)) {
if (dashstream->active_stream
&& dashstream->active_stream->cur_representation) {
/* id specifies an identifier for this Representation. The
* identifier shall be unique within a Period unless the
* Representation is functionally identically to another
* Representation in the same Period. */
if (g_strcmp0 (dashstream->active_stream->cur_representation->id,
dashstream->last_representation_id)) {
GstCaps *caps;
stream->need_header = TRUE;
GST_INFO_OBJECT (dashdemux,
"Representation changed from %s to %s - updating to bitrate %d",
GST_STR_NULL (dashstream->last_representation_id),
GST_STR_NULL (dashstream->active_stream->cur_representation->id),
dashstream->active_stream->cur_representation->bandwidth);
caps =
gst_dash_demux_get_input_caps (dashdemux,
dashstream->active_stream);
gst_adaptive_demux_stream_set_caps (stream, caps);
/* Update the stored last representation id */
g_free (dashstream->last_representation_id);
dashstream->last_representation_id =
g_strdup (dashstream->active_stream->cur_representation->id);
}
} else {
g_free (dashstream->last_representation_id);
dashstream->last_representation_id = NULL;
}
}
2014-08-26 19:45:46 +00:00
if (GST_ADAPTIVE_DEMUX_STREAM_NEED_HEADER (stream)) {
gst_adaptive_demux_stream_fragment_clear (&stream->fragment);
gst_dash_demux_stream_update_headers_info (stream);
2014-08-26 19:45:46 +00:00
}
2014-08-26 19:45:46 +00:00
gst_mpd_client_get_next_fragment (dashdemux->client, dashstream->index,
&fragment);
2012-10-17 08:02:39 +00:00
2014-08-26 19:45:46 +00:00
stream->fragment.uri = fragment.uri;
/* If mpd does not specify indexRange (i.e., null index_uri),
* sidx entries may not be available until download it */
if (isombff && dashstream->sidx_position != GST_CLOCK_TIME_NONE
&& SIDX (dashstream)->entries) {
GstSidxBoxEntry *entry = SIDX_CURRENT_ENTRY (dashstream);
stream->fragment.range_start =
dashstream->sidx_base_offset + entry->offset;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
dashstream->actual_position = stream->fragment.timestamp = entry->pts;
dashstream->current_fragment_timestamp = stream->fragment.timestamp =
entry->pts;
dashstream->current_fragment_duration = stream->fragment.duration =
entry->duration;
if (stream->demux->segment.rate < 0.0) {
stream->fragment.range_end =
stream->fragment.range_start + entry->size - 1;
dashstream->actual_position += entry->duration;
} else {
stream->fragment.range_end = fragment.range_end;
}
} else {
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
dashstream->actual_position = stream->fragment.timestamp =
fragment.timestamp;
dashstream->current_fragment_timestamp = fragment.timestamp;
dashstream->current_fragment_duration = stream->fragment.duration =
fragment.duration;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (stream->demux->segment.rate < 0.0)
dashstream->actual_position += fragment.duration;
stream->fragment.range_start =
MAX (fragment.range_start, dashstream->sidx_base_offset);
stream->fragment.range_end = fragment.range_end;
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_DEBUG_OBJECT (stream->pad, "Actual position %" GST_TIME_FORMAT,
GST_TIME_ARGS (dashstream->actual_position));
2014-08-26 19:45:46 +00:00
return GST_FLOW_OK;
}
2014-08-26 19:45:46 +00:00
return GST_FLOW_EOS;
}
2012-10-17 08:02:39 +00:00
static gint
gst_dash_demux_index_entry_search (GstSidxBoxEntry * entry, GstClockTime * ts,
gpointer user_data)
{
GstClockTime entry_ts = entry->pts + entry->duration;
if (entry_ts <= *ts)
return -1;
else if (entry->pts > *ts)
return 1;
else
return 0;
}
static GstFlowReturn
gst_dash_demux_stream_sidx_seek (GstDashDemuxStream * dashstream,
gboolean forward, GstSeekFlags flags, GstClockTime ts,
GstClockTime * final_ts)
{
GstSidxBox *sidx = SIDX (dashstream);
GstSidxBoxEntry *entry;
gint idx = sidx->entries_count;
GstFlowReturn ret = GST_FLOW_OK;
if (sidx->entries_count == 0)
return GST_FLOW_EOS;
entry =
gst_util_array_binary_search (sidx->entries, sidx->entries_count,
sizeof (GstSidxBoxEntry),
(GCompareDataFunc) gst_dash_demux_index_entry_search,
GST_SEARCH_MODE_EXACT, &ts, NULL);
/* No exact match found, nothing in our index
* This is usually a bug or broken stream, as the seeking code already
* makes sure that we're in the correct period and segment, and only need
* to find the correct place inside the segment. Allow for some rounding
* errors and inaccuracies here though */
if (!entry) {
GstSidxBoxEntry *last_entry = &sidx->entries[sidx->entries_count - 1];
GST_WARNING_OBJECT (dashstream->parent.pad, "Couldn't find SIDX entry");
if (ts < sidx->entries[0].pts
&& ts + 250 * GST_MSECOND >= sidx->entries[0].pts)
entry = &sidx->entries[0];
else if (ts >= last_entry->pts + last_entry->duration &&
ts < last_entry->pts + last_entry->duration + 250 * GST_MSECOND)
entry = last_entry;
}
if (!entry)
return GST_FLOW_EOS;
idx = entry - sidx->entries;
/* FIXME in reverse mode, if we are exactly at a fragment start it makes more
* sense to start from the end of the previous fragment */
if (!forward && idx > 0 && entry->pts == ts) {
idx--;
entry = &sidx->entries[idx];
}
/* Now entry->pts <= ts < entry->pts + entry->duration, need to adjust for
* snapping */
if ((flags & GST_SEEK_FLAG_SNAP_NEAREST) == GST_SEEK_FLAG_SNAP_NEAREST) {
if (idx + 1 < sidx->entries_count
&& sidx->entries[idx + 1].pts - ts < ts - sidx->entries[idx].pts)
idx += 1;
} else if ((forward && (flags & GST_SEEK_FLAG_SNAP_AFTER)) || (!forward
&& (flags & GST_SEEK_FLAG_SNAP_BEFORE))) {
if (idx + 1 < sidx->entries_count && entry->pts < ts)
idx += 1;
}
g_assert (sidx->entry_index < sidx->entries_count);
sidx->entry_index = idx;
dashstream->sidx_position = sidx->entries[idx].pts;
if (final_ts)
*final_ts = dashstream->sidx_position;
return ret;
}
2014-08-26 19:45:46 +00:00
static GstFlowReturn
gst_dash_demux_stream_seek (GstAdaptiveDemuxStream * stream, gboolean forward,
GstSeekFlags flags, GstClockTime ts, GstClockTime * final_ts)
2014-08-26 19:45:46 +00:00
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (stream->demux);
gint last_index, last_repeat;
gboolean is_isobmff;
last_index = dashstream->active_stream->segment_index;
last_repeat = dashstream->active_stream->segment_repeat_index;
2012-10-17 08:02:39 +00:00
if (dashstream->adapter)
gst_adapter_clear (dashstream->adapter);
dashstream->current_offset = -1;
dashstream->current_index_header_or_data = 0;
dashstream->isobmff_parser.current_fourcc = 0;
dashstream->isobmff_parser.current_start_offset = 0;
dashstream->isobmff_parser.current_size = 0;
if (dashstream->moof)
gst_isoff_moof_box_free (dashstream->moof);
dashstream->moof = NULL;
if (dashstream->moof_sync_samples)
g_array_free (dashstream->moof_sync_samples, TRUE);
dashstream->moof_sync_samples = NULL;
dashstream->current_sync_sample = -1;
dashstream->target_time = GST_CLOCK_TIME_NONE;
is_isobmff = gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client);
if (!gst_mpd_client_stream_seek (dashdemux->client, dashstream->active_stream,
forward,
is_isobmff ? (flags & (~(GST_SEEK_FLAG_SNAP_BEFORE |
GST_SEEK_FLAG_SNAP_AFTER))) : flags, ts, final_ts)) {
return GST_FLOW_EOS;
}
if (is_isobmff) {
GstClockTime period_start, offset;
period_start = gst_mpd_client_get_period_start_time (dashdemux->client);
offset =
gst_mpd_client_get_stream_presentation_offset (dashdemux->client,
dashstream->index);
if (G_UNLIKELY (ts < period_start))
ts = offset;
else
ts += offset - period_start;
if (last_index != dashstream->active_stream->segment_index ||
last_repeat != dashstream->active_stream->segment_repeat_index) {
GST_LOG_OBJECT (stream->pad,
"Segment index was changed, reset sidx parser");
gst_isoff_sidx_parser_clear (&dashstream->sidx_parser);
dashstream->sidx_base_offset = 0;
dashstream->allow_sidx = TRUE;
}
if (dashstream->sidx_parser.status == GST_ISOFF_SIDX_PARSER_FINISHED) {
if (gst_dash_demux_stream_sidx_seek (dashstream, forward, flags, ts,
final_ts) != GST_FLOW_OK) {
GST_ERROR_OBJECT (stream->pad, "Couldn't find position in sidx");
dashstream->sidx_position = GST_CLOCK_TIME_NONE;
gst_isoff_sidx_parser_clear (&dashstream->sidx_parser);
}
dashstream->pending_seek_ts = GST_CLOCK_TIME_NONE;
} else {
/* no index yet, seek when we have it */
/* FIXME - the final_ts won't be correct here */
dashstream->pending_seek_ts = ts;
}
}
stream->discont = TRUE;
2014-08-26 19:45:46 +00:00
return GST_FLOW_OK;
2013-05-08 14:13:32 +00:00
}
static gboolean
gst_dash_demux_stream_has_next_sync_sample (GstAdaptiveDemuxStream * stream)
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
if (dashstream->moof_sync_samples &&
GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (stream->demux)) {
if (stream->demux->segment.rate > 0.0) {
if (dashstream->current_sync_sample + 1 <
dashstream->moof_sync_samples->len)
return TRUE;
} else {
if (dashstream->current_sync_sample >= 1)
return TRUE;
}
}
return FALSE;
}
static gboolean
gst_dash_demux_stream_has_next_subfragment (GstAdaptiveDemuxStream * stream)
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstSidxBox *sidx = SIDX (dashstream);
if (dashstream->sidx_parser.status == GST_ISOFF_SIDX_PARSER_FINISHED) {
if (stream->demux->segment.rate > 0.0) {
if (sidx->entry_index + 1 < sidx->entries_count)
return TRUE;
} else {
if (sidx->entry_index >= 1)
return TRUE;
}
}
return FALSE;
}
static gboolean
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
gst_dash_demux_stream_advance_sync_sample (GstAdaptiveDemuxStream * stream,
GstClockTime target_time)
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
gboolean fragment_finished = FALSE;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
guint idx = -1;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (GST_CLOCK_TIME_IS_VALID (target_time)) {
GST_LOG_OBJECT (stream->pad,
"target_time:%" GST_TIME_FORMAT " fragment ts %" GST_TIME_FORMAT
" average keyframe dist: %" GST_TIME_FORMAT
" current keyframe dist: %" GST_TIME_FORMAT
" fragment duration:%" GST_TIME_FORMAT,
GST_TIME_ARGS (target_time),
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_TIME_ARGS (dashstream->current_fragment_timestamp),
GST_TIME_ARGS (dashstream->keyframe_average_distance),
GST_TIME_ARGS (dashstream->current_fragment_keyframe_distance),
GST_TIME_ARGS (stream->fragment.duration));
if (stream->demux->segment.rate > 0.0) {
idx =
(target_time -
dashstream->current_fragment_timestamp) /
dashstream->current_fragment_keyframe_distance;
/* Prevent getting stuck in a loop due to rounding errors */
if (idx == dashstream->current_sync_sample)
idx++;
} else {
GstClockTime end_time =
dashstream->current_fragment_timestamp +
dashstream->current_fragment_duration;
if (end_time < target_time) {
idx = dashstream->moof_sync_samples->len;
} else {
idx =
(end_time -
target_time) / dashstream->current_fragment_keyframe_distance;
if (idx == dashstream->moof_sync_samples->len) {
dashstream->current_sync_sample = -1;
fragment_finished = TRUE;
goto beach;
}
idx = dashstream->moof_sync_samples->len - 1 - idx;
}
/* Prevent getting stuck in a loop due to rounding errors */
if (idx == dashstream->current_sync_sample) {
if (idx == 0) {
dashstream->current_sync_sample = -1;
fragment_finished = TRUE;
goto beach;
}
idx--;
}
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
}
GST_DEBUG_OBJECT (stream->pad,
"Advancing sync sample #%d target #%d",
dashstream->current_sync_sample, idx);
if (idx != -1 && idx >= dashstream->moof_sync_samples->len) {
dashstream->current_sync_sample = -1;
fragment_finished = TRUE;
goto beach;
}
if (stream->demux->segment.rate > 0.0) {
/* Try to get the sync sample for the target time */
if (idx != -1) {
dashstream->current_sync_sample = idx;
} else {
dashstream->current_sync_sample++;
if (dashstream->current_sync_sample >= dashstream->moof_sync_samples->len) {
fragment_finished = TRUE;
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
}
} else {
if (idx != -1) {
dashstream->current_sync_sample = idx;
} else if (dashstream->current_sync_sample == -1) {
dashstream->current_sync_sample = dashstream->moof_sync_samples->len - 1;
} else if (dashstream->current_sync_sample == 0) {
dashstream->current_sync_sample = -1;
fragment_finished = TRUE;
} else {
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
dashstream->current_sync_sample--;
}
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
beach:
GST_DEBUG_OBJECT (stream->pad,
"Advancing sync sample #%d fragment_finished:%d",
dashstream->current_sync_sample, fragment_finished);
if (!fragment_finished)
stream->discont = TRUE;
return !fragment_finished;
}
static gboolean
gst_dash_demux_stream_advance_subfragment (GstAdaptiveDemuxStream * stream)
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstSidxBox *sidx = SIDX (dashstream);
gboolean fragment_finished = TRUE;
if (dashstream->sidx_parser.status == GST_ISOFF_SIDX_PARSER_FINISHED) {
if (stream->demux->segment.rate > 0.0) {
gint idx = ++sidx->entry_index;
if (idx < sidx->entries_count) {
fragment_finished = FALSE;
}
if (idx == sidx->entries_count)
dashstream->sidx_position =
sidx->entries[idx - 1].pts + sidx->entries[idx - 1].duration;
else
dashstream->sidx_position = sidx->entries[idx].pts;
} else {
gint idx = --sidx->entry_index;
if (idx >= 0) {
fragment_finished = FALSE;
dashstream->sidx_position = sidx->entries[idx].pts;
} else {
dashstream->sidx_position = GST_CLOCK_TIME_NONE;
}
}
}
GST_DEBUG_OBJECT (stream->pad, "New sidx index: %d / %d. "
"Finished fragment: %d", sidx->entry_index, sidx->entries_count,
fragment_finished);
return !fragment_finished;
}
static gboolean
gst_dash_demux_stream_has_next_fragment (GstAdaptiveDemuxStream * stream)
{
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (stream->demux);
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
if (dashstream->moof_sync_samples &&
GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (dashdemux)) {
if (gst_dash_demux_stream_has_next_sync_sample (stream))
return TRUE;
}
if (gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client)) {
if (gst_dash_demux_stream_has_next_subfragment (stream))
return TRUE;
}
return gst_mpd_client_has_next_segment (dashdemux->client,
dashstream->active_stream, stream->demux->segment.rate > 0.0);
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
/* The goal here is to figure out, once we have pushed a keyframe downstream,
* what the next ideal keyframe to download is.
*
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
* This is done based on:
* * the current internal position (i.e. actual_position)
* * the reported downstream position (QoS feedback)
* * the average keyframe download time (average_download_time)
*/
static GstClockTime
gst_dash_demux_stream_get_target_time (GstDashDemux * dashdemux,
GstAdaptiveDemuxStream * stream, GstClockTime cur_position,
GstClockTime min_skip)
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
{
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstClockTime cur_running, min_running, min_position;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GstClockTimeDiff diff;
GstClockTime ret = cur_position;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GstClockTime deadline;
GstClockTime upstream_earliest_time;
GstClockTime earliest_time = GST_CLOCK_TIME_NONE;
g_assert (min_skip > 0);
/* minimum stream position we have to skip to */
if (stream->segment.rate > 0)
min_position = cur_position + min_skip;
else if (cur_position < min_skip)
min_position = 0;
else
min_position = cur_position - min_skip;
/* Use current clock time or the QoS earliest time, whichever is further in
* the future. The QoS time is only updated on every QoS event and
* especially not if e.g. a videodecoder or converter drops a frame further
* downstream.
*
* We only use the times if we ever received a QoS event since the last
* flush, as otherwise base_time and clock might not be correct because of a
* still pre-rolling sink
*/
upstream_earliest_time =
gst_adaptive_demux_get_qos_earliest_time ((GstAdaptiveDemux *) dashdemux);
if (upstream_earliest_time != GST_CLOCK_TIME_NONE) {
GstClock *clock;
clock = gst_element_get_clock (GST_ELEMENT_CAST (dashdemux));
if (clock) {
GstClockTime base_time;
GstClockTime now_time;
base_time = gst_element_get_base_time (GST_ELEMENT_CAST (dashdemux));
now_time = gst_clock_get_time (clock);
if (now_time > base_time)
now_time -= base_time;
else
now_time = 0;
gst_object_unref (clock);
earliest_time = MAX (now_time, upstream_earliest_time);
} else {
earliest_time = upstream_earliest_time;
}
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
/* our current position in running time */
cur_running =
gst_segment_to_running_time (&stream->segment, GST_FORMAT_TIME,
cur_position);
/* the minimum position we have to skip to in running time */
min_running =
gst_segment_to_running_time (&stream->segment, GST_FORMAT_TIME,
min_position);
GST_DEBUG_OBJECT (stream->pad,
"position: current %" GST_TIME_FORMAT " min next %" GST_TIME_FORMAT,
GST_TIME_ARGS (cur_position), GST_TIME_ARGS (min_position));
GST_DEBUG_OBJECT (stream->pad,
"running time: current %" GST_TIME_FORMAT " min next %" GST_TIME_FORMAT
" earliest %" GST_TIME_FORMAT, GST_TIME_ARGS (cur_running),
GST_TIME_ARGS (min_running), GST_TIME_ARGS (earliest_time));
/* Take configured maximum video bandwidth and framerate into account */
{
GstClockTime min_run_dist, min_frame_dist, diff = 0;
guint max_fps_n, max_fps_d;
min_run_dist = min_skip / ABS (stream->segment.rate);
if (dashdemux->max_video_framerate_n != 0) {
max_fps_n = dashdemux->max_video_framerate_n;
max_fps_d = dashdemux->max_video_framerate_d;
} else {
/* more than 10 fps is not very useful if we're skipping anyway */
max_fps_n = 10;
max_fps_d = 1;
}
min_frame_dist = gst_util_uint64_scale_ceil (GST_SECOND,
max_fps_d, max_fps_n);
GST_DEBUG_OBJECT (stream->pad,
"Have max framerate %d/%d - Min dist %" GST_TIME_FORMAT
", min requested dist %" GST_TIME_FORMAT,
max_fps_n, max_fps_d,
GST_TIME_ARGS (min_run_dist), GST_TIME_ARGS (min_frame_dist));
if (min_frame_dist > min_run_dist)
diff = MAX (diff, min_frame_dist - min_run_dist);
if (dashdemux->max_bitrate != 0) {
guint64 max_bitrate = gst_util_uint64_scale_ceil (GST_SECOND,
8 * dashstream->keyframe_average_size,
dashstream->keyframe_average_distance) * ABS (stream->segment.rate);
if (max_bitrate > dashdemux->max_bitrate) {
min_frame_dist = gst_util_uint64_scale_ceil (GST_SECOND,
8 * dashstream->keyframe_average_size,
dashdemux->max_bitrate) * ABS (stream->segment.rate);
GST_DEBUG_OBJECT (stream->pad,
"Have max bitrate %u - Min dist %" GST_TIME_FORMAT
", min requested dist %" GST_TIME_FORMAT, dashdemux->max_bitrate,
GST_TIME_ARGS (min_run_dist), GST_TIME_ARGS (min_frame_dist));
if (min_frame_dist > min_run_dist)
diff = MAX (diff, min_frame_dist - min_run_dist);
}
}
if (diff > 0) {
GST_DEBUG_OBJECT (stream->pad,
"Skipping further ahead by %" GST_TIME_FORMAT, GST_TIME_ARGS (diff));
min_running += diff;
}
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (earliest_time == GST_CLOCK_TIME_NONE) {
GstClockTime run_key_dist;
run_key_dist =
dashstream->keyframe_average_distance / ABS (stream->segment.rate);
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
/* If we don't have downstream information (such as at startup or
* without live sinks), just get the next time by taking the minimum
* amount we have to skip ahead
* Except if it takes us longer to download */
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (run_key_dist > dashstream->average_download_time)
ret =
gst_segment_position_from_running_time (&stream->segment,
GST_FORMAT_TIME, min_running);
else
ret = gst_segment_position_from_running_time (&stream->segment,
GST_FORMAT_TIME,
min_running - run_key_dist + dashstream->average_download_time);
GST_DEBUG_OBJECT (stream->pad,
"Advancing to %" GST_TIME_FORMAT " (was %" GST_TIME_FORMAT ")",
GST_TIME_ARGS (ret), GST_TIME_ARGS (min_position));
goto out;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
}
/* Figure out the difference, in running time, between where we are and
* where downstream is */
diff = min_running - earliest_time;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_LOG_OBJECT (stream->pad,
"min_running %" GST_TIME_FORMAT " diff %" GST_STIME_FORMAT
" average_download %" GST_TIME_FORMAT, GST_TIME_ARGS (min_running),
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_STIME_ARGS (diff), GST_TIME_ARGS (dashstream->average_download_time));
/* Have at least 500ms or 3 keyframes safety between current position and downstream */
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
deadline = MAX (500 * GST_MSECOND, 3 * dashstream->average_download_time);
/* The furthest away we are from the current position, the least we need to advance */
if (diff < 0 || diff < deadline) {
/* Force skipping (but not more than 1s ahead) */
ret =
gst_segment_position_from_running_time (&stream->segment,
GST_FORMAT_TIME, earliest_time + MIN (deadline, GST_SECOND));
GST_DEBUG_OBJECT (stream->pad,
"MUST SKIP to at least %" GST_TIME_FORMAT " (was %" GST_TIME_FORMAT ")",
GST_TIME_ARGS (ret), GST_TIME_ARGS (min_position));
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
} else if (diff < 4 * dashstream->average_download_time) {
2019-09-02 19:08:44 +00:00
/* Go forward a bit less aggressively (and at most 1s forward) */
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
ret = gst_segment_position_from_running_time (&stream->segment,
GST_FORMAT_TIME, min_running + MIN (GST_SECOND,
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
2 * dashstream->average_download_time));
GST_DEBUG_OBJECT (stream->pad,
"MUST SKIP to at least %" GST_TIME_FORMAT " (was %" GST_TIME_FORMAT ")",
GST_TIME_ARGS (ret), GST_TIME_ARGS (min_position));
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
} else {
/* Get the next position satisfying the download time */
ret = gst_segment_position_from_running_time (&stream->segment,
GST_FORMAT_TIME, min_running);
GST_DEBUG_OBJECT (stream->pad,
"Advance to %" GST_TIME_FORMAT " (was %" GST_TIME_FORMAT ")",
GST_TIME_ARGS (ret), GST_TIME_ARGS (min_position));
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
}
out:
{
GstClockTime cur_skip =
(cur_position < ret) ? ret - cur_position : cur_position - ret;
if (dashstream->average_skip_size == 0) {
dashstream->average_skip_size = cur_skip;
} else {
dashstream->average_skip_size =
(cur_skip + 3 * dashstream->average_skip_size) / 4;
}
if (dashstream->average_skip_size >
cur_skip + dashstream->keyframe_average_distance
&& dashstream->average_skip_size > min_skip) {
if (stream->segment.rate > 0)
ret = cur_position + dashstream->average_skip_size;
else if (cur_position > dashstream->average_skip_size)
ret = cur_position - dashstream->average_skip_size;
else
ret = 0;
}
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
return ret;
}
2014-08-26 19:45:46 +00:00
static GstFlowReturn
gst_dash_demux_stream_advance_fragment (GstAdaptiveDemuxStream * stream)
2013-05-08 14:13:32 +00:00
{
2014-08-26 19:45:46 +00:00
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (stream->demux);
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GstClockTime target_time = GST_CLOCK_TIME_NONE;
GstClockTime previous_position;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GstFlowReturn ret;
GST_DEBUG_OBJECT (stream->pad, "Advance fragment");
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
/* Update download statistics */
if (dashstream->moof_sync_samples &&
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (dashdemux) &&
GST_CLOCK_TIME_IS_VALID (stream->last_download_time)) {
if (GST_CLOCK_TIME_IS_VALID (dashstream->average_download_time)) {
dashstream->average_download_time =
(3 * dashstream->average_download_time +
stream->last_download_time) / 4;
} else {
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
dashstream->average_download_time = stream->last_download_time;
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_DEBUG_OBJECT (stream->pad,
"Download time last: %" GST_TIME_FORMAT " average: %" GST_TIME_FORMAT,
GST_TIME_ARGS (stream->last_download_time),
GST_TIME_ARGS (dashstream->average_download_time));
}
previous_position = dashstream->actual_position;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
/* Update internal position */
if (GST_CLOCK_TIME_IS_VALID (dashstream->actual_position)) {
GstClockTime dur;
if (dashstream->moof_sync_samples
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (dashdemux)) {
GST_LOG_OBJECT (stream->pad, "current sync sample #%d",
dashstream->current_sync_sample);
if (dashstream->current_sync_sample == -1) {
dur = 0;
} else if (dashstream->current_sync_sample <
dashstream->moof_sync_samples->len) {
dur = dashstream->current_fragment_keyframe_distance;
} else {
if (gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client) &&
dashstream->sidx_position != GST_CLOCK_TIME_NONE
&& SIDX (dashstream)->entries) {
GstSidxBoxEntry *entry = SIDX_CURRENT_ENTRY (dashstream);
dur = entry->duration;
} else {
dur =
dashstream->current_fragment_timestamp +
dashstream->current_fragment_duration -
dashstream->actual_position;
}
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
} else if (gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client) &&
dashstream->sidx_position != GST_CLOCK_TIME_NONE
&& SIDX (dashstream)->entries) {
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GstSidxBoxEntry *entry = SIDX_CURRENT_ENTRY (dashstream);
dur = entry->duration;
} else {
dur = stream->fragment.duration;
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (dashstream->moof_sync_samples
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (dashdemux)) {
/* We just downloaded the header, we actually use the previous
* target_time now as it was not used up yet */
if (dashstream->current_sync_sample == -1)
target_time = dashstream->target_time;
else
target_time =
gst_dash_demux_stream_get_target_time (dashdemux, stream,
dashstream->actual_position, dur);
dashstream->actual_position = target_time;
} else {
/* Adjust based on direction */
if (stream->demux->segment.rate > 0.0)
dashstream->actual_position += dur;
else if (dashstream->actual_position >= dur)
dashstream->actual_position -= dur;
else
dashstream->actual_position = 0;
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_DEBUG_OBJECT (stream->pad, "Actual position %" GST_TIME_FORMAT,
GST_TIME_ARGS (dashstream->actual_position));
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
dashstream->target_time = target_time;
GST_DEBUG_OBJECT (stream->pad, "target_time: %" GST_TIME_FORMAT,
GST_TIME_ARGS (target_time));
/* If downloading only keyframes, switch to the next one or fall through */
if (dashstream->moof_sync_samples &&
GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (dashdemux)) {
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (gst_dash_demux_stream_advance_sync_sample (stream, target_time))
return GST_FLOW_OK;
}
dashstream->isobmff_parser.current_fourcc = 0;
dashstream->isobmff_parser.current_start_offset = 0;
dashstream->isobmff_parser.current_size = 0;
if (dashstream->moof)
gst_isoff_moof_box_free (dashstream->moof);
dashstream->moof = NULL;
if (dashstream->moof_sync_samples)
g_array_free (dashstream->moof_sync_samples, TRUE);
dashstream->moof_sync_samples = NULL;
dashstream->current_sync_sample = -1;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
/* Check if we just need to 'advance' to the next fragment, or if we
* need to skip by more. */
if (GST_CLOCK_TIME_IS_VALID (target_time)
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (stream->demux) &&
dashstream->active_stream->mimeType == GST_STREAM_VIDEO) {
GstClockTime actual_ts;
GstSeekFlags flags = 0;
/* Key-unit trick mode, seek to fragment containing target time
*
* We first try seeking without snapping. As above code to skip keyframes
* in the current fragment was not successful, we should go at least one
* fragment ahead. Due to rounding errors we could end up at the same
* fragment again here, in which case we retry seeking with the SNAP_AFTER
* flag.
*
* We don't always set that flag as we would then end up one further
* fragment in the future in all good cases.
*/
while (TRUE) {
ret =
gst_dash_demux_stream_seek (stream, (stream->segment.rate > 0), flags,
target_time, &actual_ts);
if (ret != GST_FLOW_OK) {
GST_WARNING_OBJECT (stream->pad, "Failed to seek to %" GST_TIME_FORMAT,
GST_TIME_ARGS (target_time));
/* Give up */
if (flags != 0)
break;
/* Retry with skipping ahead */
flags |= GST_SEEK_FLAG_SNAP_AFTER;
continue;
}
GST_DEBUG_OBJECT (stream->pad,
"Skipped to %" GST_TIME_FORMAT " (wanted %" GST_TIME_FORMAT ", was %"
GST_TIME_FORMAT ")", GST_TIME_ARGS (actual_ts),
GST_TIME_ARGS (target_time), GST_TIME_ARGS (previous_position));
if ((stream->segment.rate > 0 && actual_ts <= previous_position) ||
(stream->segment.rate < 0 && actual_ts >= previous_position)) {
/* Give up */
if (flags != 0)
break;
/* Retry with forcing skipping ahead */
flags |= GST_SEEK_FLAG_SNAP_AFTER;
continue;
}
/* All good */
break;
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
} else {
/* Normal mode, advance to the next fragment */
if (gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client)) {
if (gst_dash_demux_stream_advance_subfragment (stream))
return GST_FLOW_OK;
}
if (dashstream->adapter)
gst_adapter_clear (dashstream->adapter);
gst_isoff_sidx_parser_clear (&dashstream->sidx_parser);
dashstream->sidx_base_offset = 0;
dashstream->sidx_position = GST_CLOCK_TIME_NONE;
dashstream->allow_sidx = TRUE;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
ret = gst_mpd_client_advance_segment (dashdemux->client,
dashstream->active_stream, stream->demux->segment.rate > 0.0);
}
return ret;
2013-05-08 14:13:32 +00:00
}
2014-08-26 19:45:46 +00:00
static gboolean
gst_dash_demux_stream_select_bitrate (GstAdaptiveDemuxStream * stream,
guint64 bitrate)
2013-05-08 14:13:32 +00:00
{
GstActiveStream *active_stream = NULL;
GList *rep_list = NULL;
2013-05-08 14:13:32 +00:00
gint new_index;
GstAdaptiveDemux *base_demux = stream->demux;
2014-08-26 19:45:46 +00:00
GstDashDemux *demux = GST_DASH_DEMUX_CAST (stream->demux);
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
gboolean ret = FALSE;
2013-05-08 14:13:32 +00:00
2014-08-26 19:45:46 +00:00
active_stream = dashstream->active_stream;
if (active_stream == NULL) {
goto end;
}
/* In key-frame trick mode don't change bitrates */
if (GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (demux)) {
GST_DEBUG_OBJECT (demux, "In key-frame trick mode, not changing bitrates");
goto end;
}
/* retrieve representation list */
if (active_stream->cur_adapt_set)
rep_list = active_stream->cur_adapt_set->Representations;
2014-08-26 19:45:46 +00:00
if (!rep_list) {
goto end;
}
GST_DEBUG_OBJECT (stream->pad,
"Trying to change to bitrate: %" G_GUINT64_FORMAT, bitrate);
if (active_stream->mimeType == GST_STREAM_VIDEO && demux->max_bitrate) {
bitrate = MIN (demux->max_bitrate, bitrate);
}
/* get representation index with current max_bandwidth */
if (GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (base_demux) ||
ABS (base_demux->segment.rate) <= 1.0) {
new_index =
gst_mpd_client_get_rep_idx_with_max_bandwidth (rep_list, bitrate,
demux->max_video_width, demux->max_video_height,
demux->max_video_framerate_n, demux->max_video_framerate_d);
} else {
new_index =
gst_mpd_client_get_rep_idx_with_max_bandwidth (rep_list,
bitrate / ABS (base_demux->segment.rate), demux->max_video_width,
demux->max_video_height, demux->max_video_framerate_n,
demux->max_video_framerate_d);
}
/* if no representation has the required bandwidth, take the lowest one */
if (new_index == -1)
new_index = gst_mpd_client_get_rep_idx_with_min_bandwidth (rep_list);
if (new_index != active_stream->representation_idx) {
GstMPDRepresentationNode *rep = g_list_nth_data (rep_list, new_index);
GST_INFO_OBJECT (demux, "Changing representation idx: %d %d %u",
2014-08-26 19:45:46 +00:00
dashstream->index, new_index, rep->bandwidth);
if (gst_mpd_client_setup_representation (demux->client, active_stream, rep)) {
2014-08-26 19:45:46 +00:00
GstCaps *caps;
GST_INFO_OBJECT (demux, "Switching bitrate to %d",
active_stream->cur_representation->bandwidth);
2014-08-26 19:45:46 +00:00
caps = gst_dash_demux_get_input_caps (demux, active_stream);
gst_adaptive_demux_stream_set_caps (stream, caps);
ret = TRUE;
/* Update the stored last representation id */
g_free (dashstream->last_representation_id);
dashstream->last_representation_id =
g_strdup (active_stream->cur_representation->id);
} else {
GST_WARNING_OBJECT (demux, "Can not switch representation, aborting...");
2013-05-08 14:13:32 +00:00
}
}
if (ret) {
if (gst_mpd_client_has_isoff_ondemand_profile (demux->client)
&& SIDX (dashstream)->entries) {
/* store our current position to change to the same one in a different
* representation if needed */
if (SIDX (dashstream)->entry_index < SIDX (dashstream)->entries_count)
dashstream->sidx_position = SIDX_CURRENT_ENTRY (dashstream)->pts;
else if (SIDX (dashstream)->entry_index >=
SIDX (dashstream)->entries_count)
dashstream->sidx_position =
SIDX_ENTRY (dashstream,
SIDX (dashstream)->entries_count - 1)->pts + SIDX_ENTRY (dashstream,
SIDX (dashstream)->entries_count - 1)->duration;
else
dashstream->sidx_position = GST_CLOCK_TIME_NONE;
} else {
dashstream->sidx_position = GST_CLOCK_TIME_NONE;
}
gst_isoff_sidx_parser_clear (&dashstream->sidx_parser);
dashstream->sidx_base_offset = 0;
dashstream->allow_sidx = TRUE;
/* Reset ISOBMFF box parsing state */
dashstream->isobmff_parser.current_fourcc = 0;
dashstream->isobmff_parser.current_start_offset = 0;
dashstream->isobmff_parser.current_size = 0;
dashstream->current_offset = -1;
dashstream->current_index_header_or_data = 0;
if (dashstream->adapter)
gst_adapter_clear (dashstream->adapter);
if (dashstream->moof)
gst_isoff_moof_box_free (dashstream->moof);
dashstream->moof = NULL;
if (dashstream->moof_sync_samples)
g_array_free (dashstream->moof_sync_samples, TRUE);
dashstream->moof_sync_samples = NULL;
dashstream->current_sync_sample = -1;
dashstream->target_time = GST_CLOCK_TIME_NONE;
}
2014-08-26 19:45:46 +00:00
end:
return ret;
}
#define SEEK_UPDATES_PLAY_POSITION(r, start_type, stop_type) \
((r >= 0 && start_type != GST_SEEK_TYPE_NONE) || \
(r < 0 && stop_type != GST_SEEK_TYPE_NONE))
2014-08-26 19:45:46 +00:00
static gboolean
gst_dash_demux_seek (GstAdaptiveDemux * demux, GstEvent * seek)
2013-05-08 14:13:32 +00:00
{
2014-08-26 19:45:46 +00:00
gdouble rate;
GstFormat format;
GstSeekFlags flags;
GstSeekType start_type, stop_type;
gint64 start, stop;
GList *list;
GstClockTime current_pos, target_pos;
guint current_period;
GstStreamPeriod *period;
GList *iter, *streams = NULL;
2014-08-26 19:45:46 +00:00
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
gboolean trickmode_no_audio;
2014-08-26 19:45:46 +00:00
gst_event_parse_seek (seek, &rate, &format, &flags, &start_type, &start,
&stop_type, &stop);
if (!SEEK_UPDATES_PLAY_POSITION (rate, start_type, stop_type)) {
/* nothing to do if we don't have to update the current position */
return TRUE;
}
if (demux->segment.rate > 0.0) {
target_pos = (GstClockTime) start;
} else {
target_pos = (GstClockTime) stop;
}
2014-08-26 19:45:46 +00:00
/* select the requested Period in the Media Presentation */
if (!gst_mpd_client_setup_media_presentation (dashdemux->client, target_pos,
-1, NULL))
return FALSE;
2014-08-26 19:45:46 +00:00
current_period = 0;
for (list = g_list_first (dashdemux->client->periods); list;
list = g_list_next (list)) {
period = list->data;
current_pos = period->start;
current_period = period->number;
GST_DEBUG_OBJECT (demux, "Looking at period %u) start:%"
GST_TIME_FORMAT " - duration:%"
GST_TIME_FORMAT ") for position %" GST_TIME_FORMAT,
current_period, GST_TIME_ARGS (current_pos),
GST_TIME_ARGS (period->duration), GST_TIME_ARGS (target_pos));
2014-08-26 19:45:46 +00:00
if (current_pos <= target_pos
&& target_pos <= current_pos + period->duration) {
2014-08-26 19:45:46 +00:00
break;
}
}
2014-08-26 19:45:46 +00:00
if (list == NULL) {
GST_WARNING_OBJECT (demux, "Could not find seeked Period");
return FALSE;
}
trickmode_no_audio = !!(flags & GST_SEEK_FLAG_TRICKMODE_NO_AUDIO);
streams = demux->streams;
2014-08-26 19:45:46 +00:00
if (current_period != gst_mpd_client_get_period_index (dashdemux->client)) {
GST_DEBUG_OBJECT (demux, "Seeking to Period %d", current_period);
2014-08-26 19:45:46 +00:00
/* clean old active stream list, if any */
gst_mpd_client_active_streams_free (dashdemux->client);
dashdemux->trickmode_no_audio = trickmode_no_audio;
2014-08-26 19:45:46 +00:00
/* setup video, audio and subtitle streams, starting from the new Period */
if (!gst_mpd_client_set_period_index (dashdemux->client, current_period)
|| !gst_dash_demux_setup_all_streams (dashdemux))
return FALSE;
streams = demux->next_streams;
} else if (dashdemux->trickmode_no_audio != trickmode_no_audio) {
/* clean old active stream list, if any */
gst_mpd_client_active_streams_free (dashdemux->client);
dashdemux->trickmode_no_audio = trickmode_no_audio;
/* setup video, audio and subtitle streams, starting from the new Period */
if (!gst_dash_demux_setup_all_streams (dashdemux))
return FALSE;
streams = demux->next_streams;
}
2014-08-26 19:45:46 +00:00
/* Update the current sequence on all streams */
for (iter = streams; iter; iter = g_list_next (iter)) {
GstAdaptiveDemuxStream *stream = iter->data;
GstDashDemuxStream *dashstream = iter->data;
dashstream->average_skip_size = 0;
if (gst_dash_demux_stream_seek (stream, rate >= 0, 0, target_pos,
NULL) != GST_FLOW_OK)
return FALSE;
2013-05-08 14:13:32 +00:00
}
2014-08-26 19:45:46 +00:00
return TRUE;
2013-05-08 14:13:32 +00:00
}
2014-08-26 19:45:46 +00:00
static gint64
gst_dash_demux_get_manifest_update_interval (GstAdaptiveDemux * demux)
{
2014-08-26 19:45:46 +00:00
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
return MIN (dashdemux->client->mpd_root_node->minimumUpdatePeriod * 1000,
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
SLOW_CLOCK_UPDATE_INTERVAL);
}
static GstFlowReturn
gst_dash_demux_update_manifest_data (GstAdaptiveDemux * demux,
GstBuffer * buffer)
{
2014-08-26 19:45:46 +00:00
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
GstMPDClient *new_client = NULL;
2014-08-26 19:45:46 +00:00
GstMapInfo mapinfo;
2014-08-26 19:45:46 +00:00
GST_DEBUG_OBJECT (demux, "Updating manifest file from URL");
2014-08-26 19:45:46 +00:00
/* parse the manifest file */
new_client = gst_mpd_client_new ();
gst_mpd_client_set_uri_downloader (new_client, demux->downloader);
new_client->mpd_uri = g_strdup (demux->manifest_uri);
new_client->mpd_base_uri = g_strdup (demux->manifest_base_uri);
2014-08-26 19:45:46 +00:00
gst_buffer_map (buffer, &mapinfo, GST_MAP_READ);
if (gst_mpd_client_parse (new_client, (gchar *) mapinfo.data, mapinfo.size)) {
2014-08-26 19:45:46 +00:00
const gchar *period_id;
guint period_idx;
GList *iter;
GList *streams_iter;
GList *streams;
2014-08-26 19:45:46 +00:00
/* prepare the new manifest and try to transfer the stream position
* status from the old manifest client */
2014-08-26 19:45:46 +00:00
GST_DEBUG_OBJECT (demux, "Updating manifest");
2014-08-26 19:45:46 +00:00
period_id = gst_mpd_client_get_period_id (dashdemux->client);
period_idx = gst_mpd_client_get_period_index (dashdemux->client);
2014-08-26 19:45:46 +00:00
/* setup video, audio and subtitle streams, starting from current Period */
if (!gst_mpd_client_setup_media_presentation (new_client, -1,
(period_id ? -1 : period_idx), period_id)) {
2014-08-26 19:45:46 +00:00
/* TODO */
}
if (period_id) {
2014-08-26 19:45:46 +00:00
if (!gst_mpd_client_set_period_id (new_client, period_id)) {
GST_DEBUG_OBJECT (demux, "Error setting up the updated manifest file");
gst_mpd_client_free (new_client);
gst_buffer_unmap (buffer, &mapinfo);
2014-08-26 19:45:46 +00:00
return GST_FLOW_EOS;
}
} else {
2014-08-26 19:45:46 +00:00
if (!gst_mpd_client_set_period_index (new_client, period_idx)) {
GST_DEBUG_OBJECT (demux, "Error setting up the updated manifest file");
gst_mpd_client_free (new_client);
gst_buffer_unmap (buffer, &mapinfo);
2014-08-26 19:45:46 +00:00
return GST_FLOW_EOS;
}
}
2014-08-26 19:45:46 +00:00
if (!gst_dash_demux_setup_mpdparser_streams (dashdemux, new_client)) {
GST_ERROR_OBJECT (demux, "Failed to setup streams on manifest " "update");
gst_mpd_client_free (new_client);
gst_buffer_unmap (buffer, &mapinfo);
2014-08-26 19:45:46 +00:00
return GST_FLOW_ERROR;
}
/* If no pads have been exposed yet, need to use those */
streams = NULL;
if (demux->streams == NULL) {
if (demux->prepared_streams) {
streams = demux->prepared_streams;
}
} else {
streams = demux->streams;
}
/* update the streams to preserve the current representation if there is one,
* and to play from the next segment */
for (iter = streams, streams_iter = new_client->active_streams;
iter && streams_iter;
iter = g_list_next (iter), streams_iter = g_list_next (streams_iter)) {
2014-08-26 19:45:46 +00:00
GstDashDemuxStream *demux_stream = iter->data;
GstActiveStream *new_stream = streams_iter->data;
2014-08-26 19:45:46 +00:00
GstClockTime ts;
2014-08-26 19:45:46 +00:00
if (!new_stream) {
GST_DEBUG_OBJECT (demux,
"Stream of index %d is missing from manifest update",
demux_stream->index);
gst_mpd_client_free (new_client);
gst_buffer_unmap (buffer, &mapinfo);
2014-08-26 19:45:46 +00:00
return GST_FLOW_EOS;
}
if (new_stream->cur_adapt_set
&& demux_stream->last_representation_id != NULL) {
GList *rep_list = new_stream->cur_adapt_set->Representations;
GstMPDRepresentationNode *rep_node =
gst_mpd_client_get_representation_with_id (rep_list,
demux_stream->last_representation_id);
if (rep_node != NULL) {
if (gst_mpd_client_setup_representation (new_client, new_stream,
rep_node)) {
GST_DEBUG_OBJECT (GST_ADAPTIVE_DEMUX_STREAM_PAD (demux_stream),
"Found and set up matching representation %s in new manifest",
demux_stream->last_representation_id);
} else {
GST_ERROR_OBJECT (GST_ADAPTIVE_DEMUX_STREAM_PAD (demux_stream),
"Failed to set up representation %s in new manifest",
demux_stream->last_representation_id);
gst_mpd_client_free (new_client);
gst_buffer_unmap (buffer, &mapinfo);
return GST_FLOW_EOS;
}
} else {
/* If we failed to find the current representation,
* then update_fragment_info() will reconfigure to the
* new settings after the current download finishes */
GST_WARNING_OBJECT (GST_ADAPTIVE_DEMUX_STREAM_PAD (demux_stream),
"Failed to find representation %s in new manifest",
demux_stream->last_representation_id);
}
}
2014-08-26 19:45:46 +00:00
if (gst_mpd_client_get_next_fragment_timestamp (dashdemux->client,
demux_stream->index, &ts)
|| gst_mpd_client_get_last_fragment_timestamp_end (dashdemux->client,
2014-08-26 19:45:46 +00:00
demux_stream->index, &ts)) {
/* Due to rounding when doing the timescale conversions it might happen
* that the ts falls back to a previous segment, leading the same data
* to be downloaded twice. We try to work around this by always adding
* 10 microseconds to get back to the correct segment. The errors are
* usually on the order of nanoseconds so it should be enough.
*/
/* _get_next_fragment_timestamp() returned relative timestamp to
* corresponding period start, but _client_stream_seek expects absolute
* MPD time. */
ts += gst_mpd_client_get_period_start_time (dashdemux->client);
GST_DEBUG_OBJECT (GST_ADAPTIVE_DEMUX_STREAM_PAD (demux_stream),
"Current position: %" GST_TIME_FORMAT ", updating to %"
GST_TIME_FORMAT, GST_TIME_ARGS (ts),
GST_TIME_ARGS (ts + (10 * GST_USECOND)));
ts += 10 * GST_USECOND;
gst_mpd_client_stream_seek (new_client, new_stream,
demux->segment.rate >= 0, 0, ts, NULL);
}
demux_stream->active_stream = new_stream;
}
2014-08-26 19:45:46 +00:00
gst_mpd_client_free (dashdemux->client);
dashdemux->client = new_client;
2014-08-26 19:45:46 +00:00
GST_DEBUG_OBJECT (demux, "Manifest file successfully updated");
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
if (dashdemux->clock_drift) {
gst_dash_demux_poll_clock_drift (dashdemux);
}
} else {
2014-08-26 19:45:46 +00:00
/* In most cases, this will happen if we set a wrong url in the
* source element and we have received the 404 HTML response instead of
* the manifest */
GST_WARNING_OBJECT (demux, "Error parsing the manifest.");
gst_mpd_client_free (new_client);
2014-08-26 19:45:46 +00:00
gst_buffer_unmap (buffer, &mapinfo);
return GST_FLOW_ERROR;
}
2014-08-26 19:45:46 +00:00
gst_buffer_unmap (buffer, &mapinfo);
2014-08-26 19:45:46 +00:00
return GST_FLOW_OK;
}
2014-08-26 19:45:46 +00:00
static gint64
gst_dash_demux_stream_get_fragment_waiting_time (GstAdaptiveDemuxStream *
stream)
{
2014-08-26 19:45:46 +00:00
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (stream->demux);
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
GstDateTime *segmentAvailability;
2014-08-26 19:45:46 +00:00
GstActiveStream *active_stream = dashstream->active_stream;
segmentAvailability =
gst_mpd_client_get_next_segment_availability_start_time
(dashdemux->client, active_stream);
if (segmentAvailability) {
2014-08-26 19:45:46 +00:00
gint64 diff;
GstDateTime *cur_time;
cur_time =
gst_date_time_new_from_g_date_time
(gst_adaptive_demux_get_client_now_utc (GST_ADAPTIVE_DEMUX_CAST
(dashdemux)));
diff =
gst_mpd_client_calculate_time_difference (cur_time,
segmentAvailability);
gst_date_time_unref (segmentAvailability);
2014-08-26 19:45:46 +00:00
gst_date_time_unref (cur_time);
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
/* subtract the server's clock drift, so that if the server's
time is behind our idea of UTC, we need to sleep for longer
before requesting a fragment */
return diff -
gst_dash_demux_get_clock_compensation (dashdemux) * GST_USECOND;
}
2014-08-26 19:45:46 +00:00
return 0;
}
2014-08-26 19:45:46 +00:00
static gboolean
gst_dash_demux_has_next_period (GstAdaptiveDemux * demux)
{
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
2014-08-26 19:45:46 +00:00
if (demux->segment.rate >= 0)
return gst_mpd_client_has_next_period (dashdemux->client);
else
2014-08-26 19:45:46 +00:00
return gst_mpd_client_has_previous_period (dashdemux->client);
}
2014-08-26 19:45:46 +00:00
static void
gst_dash_demux_advance_period (GstAdaptiveDemux * demux)
2013-05-08 14:13:32 +00:00
{
2014-08-26 19:45:46 +00:00
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
2014-08-26 19:45:46 +00:00
if (demux->segment.rate >= 0) {
if (!gst_mpd_client_set_period_index (dashdemux->client,
gst_mpd_client_get_period_index (dashdemux->client) + 1)) {
/* TODO error */
return;
}
} else {
if (!gst_mpd_client_set_period_index (dashdemux->client,
gst_mpd_client_get_period_index (dashdemux->client) - 1)) {
/* TODO error */
return;
}
}
2014-08-26 19:45:46 +00:00
gst_dash_demux_setup_all_streams (dashdemux);
gst_mpd_client_seek_to_first_segment (dashdemux->client);
}
static GstBuffer *
_gst_buffer_split (GstBuffer * buffer, gint offset, gsize size)
{
GstBuffer *newbuf = gst_buffer_copy_region (buffer,
GST_BUFFER_COPY_FLAGS | GST_BUFFER_COPY_TIMESTAMPS | GST_BUFFER_COPY_META
| GST_BUFFER_COPY_MEMORY, offset, size == -1 ? size : size - offset);
gst_buffer_resize (buffer, 0, offset);
return newbuf;
}
static gboolean
gst_dash_demux_stream_fragment_start (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream)
{
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_LOG_OBJECT (stream->pad, "Actual position %" GST_TIME_FORMAT,
GST_TIME_ARGS (dashstream->actual_position));
dashstream->current_index_header_or_data = 0;
dashstream->current_offset = -1;
/* We need to mark every first buffer of a key unit as discont,
* and also every first buffer of a moov and moof. This ensures
* that qtdemux takes note of our buffer offsets for each of those
* buffers instead of keeping track of them itself from the first
* buffer. We need offsets to be consistent between moof and mdat
*/
if (dashstream->is_isobmff && dashdemux->allow_trickmode_key_units
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (demux)
&& dashstream->active_stream->mimeType == GST_STREAM_VIDEO)
stream->discont = TRUE;
return TRUE;
}
static GstFlowReturn
gst_dash_demux_stream_fragment_finished (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream)
{
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
/* We need to mark every first buffer of a key unit as discont,
* and also every first buffer of a moov and moof. This ensures
* that qtdemux takes note of our buffer offsets for each of those
* buffers instead of keeping track of them itself from the first
* buffer. We need offsets to be consistent between moof and mdat
*/
if (dashstream->is_isobmff && dashdemux->allow_trickmode_key_units
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (demux)
&& dashstream->active_stream->mimeType == GST_STREAM_VIDEO)
stream->discont = TRUE;
/* Only handle fragment advancing specifically for SIDX if we're not
* in key unit mode */
if (!(dashstream->moof_sync_samples
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (dashdemux))
&& gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client)
&& dashstream->sidx_parser.status == GST_ISOFF_SIDX_PARSER_FINISHED) {
/* fragment is advanced on data_received when byte limits are reached */
if (dashstream->pending_seek_ts != GST_CLOCK_TIME_NONE) {
if (SIDX (dashstream)->entry_index < SIDX (dashstream)->entries_count)
return GST_FLOW_OK;
} else if (gst_dash_demux_stream_has_next_subfragment (stream)) {
return GST_FLOW_OK;
}
}
if (G_UNLIKELY (stream->downloading_header || stream->downloading_index))
return GST_FLOW_OK;
return gst_adaptive_demux_stream_advance_fragment (demux, stream,
stream->fragment.duration);
}
static gboolean
gst_dash_demux_need_another_chunk (GstAdaptiveDemuxStream * stream)
{
GstDashDemux *dashdemux = (GstDashDemux *) stream->demux;
GstDashDemuxStream *dashstream = (GstDashDemuxStream *) stream;
/* We're chunked downloading for ISOBMFF in KEY_UNITS mode for the actual
* fragment until we parsed the moof and arrived at the mdat. 8192 is a
* random guess for the moof size
*/
if (dashstream->is_isobmff
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (stream->demux)
&& dashstream->active_stream->mimeType == GST_STREAM_VIDEO
&& !stream->downloading_header && !stream->downloading_index
&& dashdemux->allow_trickmode_key_units) {
if (dashstream->isobmff_parser.current_fourcc != GST_ISOFF_FOURCC_MDAT) {
/* Need to download the moof first to know anything */
stream->fragment.chunk_size = 8192;
/* Do we have the first fourcc already or are we in the middle */
if (dashstream->isobmff_parser.current_fourcc == 0) {
stream->fragment.chunk_size += dashstream->moof_average_size;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (dashstream->first_sync_sample_always_after_moof) {
gboolean first = FALSE;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
/* Check if we'll really need that first sample */
if (GST_CLOCK_TIME_IS_VALID (dashstream->target_time)) {
first =
((dashstream->target_time -
dashstream->current_fragment_timestamp) /
dashstream->keyframe_average_distance) == 0 ? TRUE : FALSE;
} else if (stream->segment.rate > 0) {
first = TRUE;
}
if (first)
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
stream->fragment.chunk_size += dashstream->keyframe_average_size;
}
}
if (gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client) &&
dashstream->sidx_parser.sidx.entries) {
guint64 sidx_start_offset =
dashstream->sidx_base_offset +
SIDX_CURRENT_ENTRY (dashstream)->offset;
guint64 sidx_end_offset =
sidx_start_offset + SIDX_CURRENT_ENTRY (dashstream)->size;
guint64 downloaded_end_offset;
if (dashstream->current_offset == GST_CLOCK_TIME_NONE) {
downloaded_end_offset = sidx_start_offset;
} else {
downloaded_end_offset =
dashstream->current_offset +
gst_adapter_available (dashstream->adapter);
}
downloaded_end_offset = MAX (downloaded_end_offset, sidx_start_offset);
if (stream->fragment.chunk_size +
downloaded_end_offset > sidx_end_offset) {
stream->fragment.chunk_size = sidx_end_offset - downloaded_end_offset;
}
}
} else if (dashstream->moof && dashstream->moof_sync_samples) {
/* Have the moof, either we're done now or we want to download the
* directly following sync sample */
if (dashstream->first_sync_sample_after_moof
&& dashstream->current_sync_sample == 0) {
GstDashStreamSyncSample *sync_sample =
&g_array_index (dashstream->moof_sync_samples,
GstDashStreamSyncSample, 0);
guint64 end_offset = sync_sample->end_offset + 1;
guint64 downloaded_end_offset;
downloaded_end_offset =
dashstream->current_offset +
gst_adapter_available (dashstream->adapter);
if (gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client) &&
dashstream->sidx_parser.sidx.entries) {
guint64 sidx_end_offset =
dashstream->sidx_base_offset +
SIDX_CURRENT_ENTRY (dashstream)->offset +
SIDX_CURRENT_ENTRY (dashstream)->size;
if (end_offset > sidx_end_offset) {
end_offset = sidx_end_offset;
}
}
if (downloaded_end_offset < end_offset) {
stream->fragment.chunk_size = end_offset - downloaded_end_offset;
} else {
stream->fragment.chunk_size = 0;
}
} else {
stream->fragment.chunk_size = 0;
}
} else {
/* Have moof but can't do key-units mode, just download until the end */
stream->fragment.chunk_size = -1;
}
} else {
/* We might've decided that we can't allow key-unit only
* trickmodes while doing chunked downloading. In that case
* just download from here to the end now */
if (dashstream->moof
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (stream->demux)) {
stream->fragment.chunk_size = -1;
} else {
stream->fragment.chunk_size = 0;
}
}
return stream->fragment.chunk_size != 0;
}
static GstFlowReturn
gst_dash_demux_parse_isobmff (GstAdaptiveDemux * demux,
GstDashDemuxStream * dash_stream, gboolean * sidx_seek_needed)
{
GstAdaptiveDemuxStream *stream = (GstAdaptiveDemuxStream *) dash_stream;
GstDashDemux *dashdemux = GST_DASH_DEMUX_CAST (demux);
gsize available;
GstBuffer *buffer;
GstMapInfo map;
GstByteReader reader;
guint32 fourcc;
guint header_size;
guint64 size, buffer_offset;
*sidx_seek_needed = FALSE;
/* This must not be called when we're in the mdat. We only look at the mdat
* header and then stop parsing the boxes as we're only interested in the
* metadata! Handling mdat is the job of the surrounding code, as well as
* stopping or starting the next fragment when mdat is over (=> sidx)
*/
g_assert (dash_stream->isobmff_parser.current_fourcc !=
GST_ISOFF_FOURCC_MDAT);
available = gst_adapter_available (dash_stream->adapter);
buffer = gst_adapter_take_buffer (dash_stream->adapter, available);
buffer_offset = dash_stream->current_offset;
/* Always at the start of a box here */
g_assert (dash_stream->isobmff_parser.current_size == 0);
/* At the start of a box => Parse it */
gst_buffer_map (buffer, &map, GST_MAP_READ);
gst_byte_reader_init (&reader, map.data, map.size);
/* While there are more boxes left to parse ... */
dash_stream->isobmff_parser.current_start_offset = buffer_offset;
do {
dash_stream->isobmff_parser.current_fourcc = 0;
dash_stream->isobmff_parser.current_size = 0;
if (!gst_isoff_parse_box_header (&reader, &fourcc, NULL, &header_size,
&size)) {
break;
}
dash_stream->isobmff_parser.current_fourcc = fourcc;
if (size == 0) {
/* We assume this is mdat, anything else with "size until end"
* does not seem to make sense */
g_assert (dash_stream->isobmff_parser.current_fourcc ==
GST_ISOFF_FOURCC_MDAT);
dash_stream->isobmff_parser.current_size = -1;
break;
}
dash_stream->isobmff_parser.current_size = size;
/* Do we have the complete box or are at MDAT */
if (gst_byte_reader_get_remaining (&reader) < size - header_size ||
dash_stream->isobmff_parser.current_fourcc == GST_ISOFF_FOURCC_MDAT) {
/* Reset byte reader to the beginning of the box */
gst_byte_reader_set_pos (&reader,
gst_byte_reader_get_pos (&reader) - header_size);
break;
}
GST_LOG_OBJECT (stream->pad,
"box %" GST_FOURCC_FORMAT " at offset %" G_GUINT64_FORMAT " size %"
G_GUINT64_FORMAT, GST_FOURCC_ARGS (fourcc),
dash_stream->isobmff_parser.current_start_offset, size);
if (dash_stream->isobmff_parser.current_fourcc == GST_ISOFF_FOURCC_MOOF) {
GstByteReader sub_reader;
/* Only allow SIDX before the very first moof */
dash_stream->allow_sidx = FALSE;
g_assert (dash_stream->moof == NULL);
g_assert (dash_stream->moof_sync_samples == NULL);
gst_byte_reader_get_sub_reader (&reader, &sub_reader, size - header_size);
dash_stream->moof = gst_isoff_moof_box_parse (&sub_reader);
dash_stream->moof_offset =
dash_stream->isobmff_parser.current_start_offset;
dash_stream->moof_size = size;
dash_stream->current_sync_sample = -1;
if (dash_stream->moof_average_size) {
if (dash_stream->moof_average_size < size)
dash_stream->moof_average_size =
(size * 3 + dash_stream->moof_average_size) / 4;
else
dash_stream->moof_average_size =
(size + dash_stream->moof_average_size + 3) / 4;
} else {
dash_stream->moof_average_size = size;
}
} else if (dash_stream->isobmff_parser.current_fourcc ==
GST_ISOFF_FOURCC_SIDX &&
gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client) &&
dash_stream->allow_sidx) {
GstByteReader sub_reader;
GstIsoffParserResult res;
guint dummy;
dash_stream->sidx_base_offset =
dash_stream->isobmff_parser.current_start_offset + size;
dash_stream->allow_sidx = FALSE;
gst_byte_reader_get_sub_reader (&reader, &sub_reader, size - header_size);
res =
gst_isoff_sidx_parser_parse (&dash_stream->sidx_parser, &sub_reader,
&dummy);
if (res == GST_ISOFF_PARSER_DONE) {
guint64 first_offset = dash_stream->sidx_parser.sidx.first_offset;
GstSidxBox *sidx = SIDX (dash_stream);
guint i;
if (first_offset) {
GST_LOG_OBJECT (stream->pad,
"non-zero sidx first offset %" G_GUINT64_FORMAT, first_offset);
dash_stream->sidx_base_offset += first_offset;
}
for (i = 0; i < sidx->entries_count; i++) {
GstSidxBoxEntry *entry = &sidx->entries[i];
if (entry->ref_type != 0) {
GST_FIXME_OBJECT (stream->pad, "SIDX ref_type 1 not supported yet");
dash_stream->sidx_position = GST_CLOCK_TIME_NONE;
gst_isoff_sidx_parser_clear (&dash_stream->sidx_parser);
break;
}
}
/* We might've cleared the index above */
if (sidx->entries_count > 0) {
if (GST_CLOCK_TIME_IS_VALID (dash_stream->pending_seek_ts)) {
/* FIXME, preserve seek flags */
if (gst_dash_demux_stream_sidx_seek (dash_stream,
demux->segment.rate >= 0, 0, dash_stream->pending_seek_ts,
NULL) != GST_FLOW_OK) {
GST_ERROR_OBJECT (stream->pad, "Couldn't find position in sidx");
dash_stream->sidx_position = GST_CLOCK_TIME_NONE;
gst_isoff_sidx_parser_clear (&dash_stream->sidx_parser);
}
dash_stream->pending_seek_ts = GST_CLOCK_TIME_NONE;
} else {
if (dash_stream->sidx_position == GST_CLOCK_TIME_NONE) {
SIDX (dash_stream)->entry_index = 0;
} else {
if (gst_dash_demux_stream_sidx_seek (dash_stream,
demux->segment.rate >= 0, GST_SEEK_FLAG_SNAP_BEFORE,
dash_stream->sidx_position, NULL) != GST_FLOW_OK) {
GST_ERROR_OBJECT (stream->pad,
"Couldn't find position in sidx");
dash_stream->sidx_position = GST_CLOCK_TIME_NONE;
gst_isoff_sidx_parser_clear (&dash_stream->sidx_parser);
}
}
dash_stream->sidx_position =
SIDX (dash_stream)->entries[SIDX (dash_stream)->entry_index].
pts;
}
}
if (dash_stream->sidx_parser.status == GST_ISOFF_SIDX_PARSER_FINISHED &&
SIDX (dash_stream)->entry_index != 0) {
/* Need to jump to the requested SIDX entry. Push everything up to
* the SIDX box below and let the caller handle everything else */
*sidx_seek_needed = TRUE;
break;
}
}
} else {
gst_byte_reader_skip (&reader, size - header_size);
}
dash_stream->isobmff_parser.current_fourcc = 0;
dash_stream->isobmff_parser.current_start_offset += size;
dash_stream->isobmff_parser.current_size = 0;
} while (gst_byte_reader_get_remaining (&reader) > 0);
gst_buffer_unmap (buffer, &map);
/* mdat? Push all we have and wait for it to be over */
if (dash_stream->isobmff_parser.current_fourcc == GST_ISOFF_FOURCC_MDAT) {
GstBuffer *pending;
GST_LOG_OBJECT (stream->pad,
"box %" GST_FOURCC_FORMAT " at offset %" G_GUINT64_FORMAT " size %"
G_GUINT64_FORMAT, GST_FOURCC_ARGS (fourcc),
dash_stream->isobmff_parser.current_start_offset,
dash_stream->isobmff_parser.current_size);
/* At mdat. Move the start of the mdat to the adapter and have everything
* else be pushed. We parsed all header boxes at this point and are not
* supposed to be called again until the next moof */
pending = _gst_buffer_split (buffer, gst_byte_reader_get_pos (&reader), -1);
gst_adapter_push (dash_stream->adapter, pending);
dash_stream->current_offset += gst_byte_reader_get_pos (&reader);
dash_stream->isobmff_parser.current_size = 0;
GST_BUFFER_OFFSET (buffer) = buffer_offset;
GST_BUFFER_OFFSET_END (buffer) =
buffer_offset + gst_buffer_get_size (buffer);
return gst_adaptive_demux_stream_push_buffer (stream, buffer);
} else if (gst_byte_reader_get_pos (&reader) != 0) {
GstBuffer *pending;
/* Multiple complete boxes and no mdat? Push them and keep the remainder,
* which is the start of the next box if any remainder */
pending = _gst_buffer_split (buffer, gst_byte_reader_get_pos (&reader), -1);
gst_adapter_push (dash_stream->adapter, pending);
dash_stream->current_offset += gst_byte_reader_get_pos (&reader);
dash_stream->isobmff_parser.current_size = 0;
GST_BUFFER_OFFSET (buffer) = buffer_offset;
GST_BUFFER_OFFSET_END (buffer) =
buffer_offset + gst_buffer_get_size (buffer);
return gst_adaptive_demux_stream_push_buffer (stream, buffer);
}
/* Not even a single complete, non-mdat box, wait */
dash_stream->isobmff_parser.current_size = 0;
gst_adapter_push (dash_stream->adapter, buffer);
return GST_FLOW_OK;
}
static gboolean
gst_dash_demux_find_sync_samples (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream)
{
GstDashDemux *dashdemux = (GstDashDemux *) stream->demux;
GstDashDemuxStream *dash_stream = (GstDashDemuxStream *) stream;
guint i;
guint32 track_id = 0;
guint64 prev_traf_end;
gboolean trex_sample_flags = FALSE;
if (!dash_stream->moof) {
dashdemux->allow_trickmode_key_units = FALSE;
return FALSE;
}
dash_stream->current_sync_sample = -1;
dash_stream->moof_sync_samples =
g_array_new (FALSE, FALSE, sizeof (GstDashStreamSyncSample));
prev_traf_end = dash_stream->moof_offset;
/* generate table of keyframes and offsets */
for (i = 0; i < dash_stream->moof->traf->len; i++) {
GstTrafBox *traf = &g_array_index (dash_stream->moof->traf, GstTrafBox, i);
guint64 traf_offset = 0, prev_trun_end;
guint j;
if (i == 0) {
track_id = traf->tfhd.track_id;
} else if (track_id != traf->tfhd.track_id) {
GST_ERROR_OBJECT (stream->pad,
"moof with trafs of different track ids (%u != %u)", track_id,
traf->tfhd.track_id);
g_array_free (dash_stream->moof_sync_samples, TRUE);
dash_stream->moof_sync_samples = NULL;
dashdemux->allow_trickmode_key_units = FALSE;
return FALSE;
}
if (traf->tfhd.flags & GST_TFHD_FLAGS_BASE_DATA_OFFSET_PRESENT) {
traf_offset = traf->tfhd.base_data_offset;
} else if (traf->tfhd.flags & GST_TFHD_FLAGS_DEFAULT_BASE_IS_MOOF) {
traf_offset = dash_stream->moof_offset;
} else {
traf_offset = prev_traf_end;
}
prev_trun_end = traf_offset;
for (j = 0; j < traf->trun->len; j++) {
GstTrunBox *trun = &g_array_index (traf->trun, GstTrunBox, j);
guint64 trun_offset, prev_sample_end;
guint k;
if (trun->flags & GST_TRUN_FLAGS_DATA_OFFSET_PRESENT) {
trun_offset = traf_offset + trun->data_offset;
} else {
trun_offset = prev_trun_end;
}
prev_sample_end = trun_offset;
for (k = 0; k < trun->samples->len; k++) {
GstTrunSample *sample =
&g_array_index (trun->samples, GstTrunSample, k);
guint64 sample_offset;
guint32 sample_flags;
#if 0
guint32 sample_duration;
#endif
sample_offset = prev_sample_end;
if (trun->flags & GST_TRUN_FLAGS_SAMPLE_FLAGS_PRESENT) {
sample_flags = sample->sample_flags;
} else if ((trun->flags & GST_TRUN_FLAGS_FIRST_SAMPLE_FLAGS_PRESENT)
&& k == 0) {
sample_flags = trun->first_sample_flags;
} else if (traf->tfhd.
flags & GST_TFHD_FLAGS_DEFAULT_SAMPLE_FLAGS_PRESENT) {
sample_flags = traf->tfhd.default_sample_flags;
} else {
trex_sample_flags = TRUE;
continue;
}
#if 0
if (trun->flags & GST_TRUN_FLAGS_SAMPLE_DURATION_PRESENT) {
sample_duration = sample->sample_duration;
} else if (traf->tfhd.
flags & GST_TFHD_FLAGS_DEFAULT_SAMPLE_DURATION_PRESENT) {
sample_duration = traf->tfhd.default_sample_duration;
} else {
GST_FIXME_OBJECT (stream->pad,
"Sample duration given by trex - can't download only keyframes");
g_array_free (dash_stream->moof_sync_samples, TRUE);
dash_stream->moof_sync_samples = NULL;
return FALSE;
}
#endif
if (trun->flags & GST_TRUN_FLAGS_SAMPLE_SIZE_PRESENT) {
prev_sample_end += sample->sample_size;
} else if (traf->tfhd.
flags & GST_TFHD_FLAGS_DEFAULT_SAMPLE_SIZE_PRESENT) {
prev_sample_end += traf->tfhd.default_sample_size;
} else {
GST_FIXME_OBJECT (stream->pad,
"Sample size given by trex - can't download only keyframes");
g_array_free (dash_stream->moof_sync_samples, TRUE);
dash_stream->moof_sync_samples = NULL;
dashdemux->allow_trickmode_key_units = FALSE;
return FALSE;
}
/* Non-non-sync sample aka sync sample */
if (!GST_ISOFF_SAMPLE_FLAGS_SAMPLE_IS_NON_SYNC_SAMPLE (sample_flags) ||
GST_ISOFF_SAMPLE_FLAGS_SAMPLE_DEPENDS_ON (sample_flags) == 2) {
GstDashStreamSyncSample sync_sample =
{ sample_offset, prev_sample_end - 1 };
/* TODO: need timestamps so we can decide to download or not */
g_array_append_val (dash_stream->moof_sync_samples, sync_sample);
}
}
prev_trun_end = prev_sample_end;
}
prev_traf_end = prev_trun_end;
}
if (trex_sample_flags) {
if (dash_stream->moof_sync_samples->len > 0) {
GST_LOG_OBJECT (stream->pad,
"Some sample flags given by trex but still found sync samples");
} else {
GST_FIXME_OBJECT (stream->pad,
"Sample flags given by trex - can't download only keyframes");
g_array_free (dash_stream->moof_sync_samples, TRUE);
dash_stream->moof_sync_samples = NULL;
dashdemux->allow_trickmode_key_units = FALSE;
return FALSE;
}
}
if (dash_stream->moof_sync_samples->len == 0) {
GST_LOG_OBJECT (stream->pad, "No sync samples found in fragment");
g_array_free (dash_stream->moof_sync_samples, TRUE);
dash_stream->moof_sync_samples = NULL;
dashdemux->allow_trickmode_key_units = FALSE;
return FALSE;
}
{
GstDashStreamSyncSample *sync_sample;
guint i;
guint size;
GstClockTime current_keyframe_distance;
for (i = 0; i < dash_stream->moof_sync_samples->len; i++) {
sync_sample =
&g_array_index (dash_stream->moof_sync_samples,
GstDashStreamSyncSample, i);
size = sync_sample->end_offset + 1 - sync_sample->start_offset;
if (dash_stream->keyframe_average_size) {
/* Over-estimate the keyframe size */
if (dash_stream->keyframe_average_size < size)
dash_stream->keyframe_average_size =
(size * 3 + dash_stream->keyframe_average_size) / 4;
else
dash_stream->keyframe_average_size =
(size + dash_stream->keyframe_average_size * 3) / 4;
} else {
dash_stream->keyframe_average_size = size;
}
if (i == 0) {
if (dash_stream->moof_offset + dash_stream->moof_size + 8 <
sync_sample->start_offset) {
dash_stream->first_sync_sample_after_moof = FALSE;
dash_stream->first_sync_sample_always_after_moof = FALSE;
} else {
dash_stream->first_sync_sample_after_moof =
(dash_stream->moof_sync_samples->len == 1
|| demux->segment.rate > 0.0);
}
}
}
g_assert (stream->fragment.duration != 0);
g_assert (stream->fragment.duration != GST_CLOCK_TIME_NONE);
if (gst_mpd_client_has_isoff_ondemand_profile (dashdemux->client)
&& dash_stream->sidx_position != GST_CLOCK_TIME_NONE
&& SIDX (dash_stream)->entries) {
GstSidxBoxEntry *entry = SIDX_CURRENT_ENTRY (dash_stream);
current_keyframe_distance =
entry->duration / dash_stream->moof_sync_samples->len;
} else {
current_keyframe_distance =
stream->fragment.duration / dash_stream->moof_sync_samples->len;
}
dash_stream->current_fragment_keyframe_distance = current_keyframe_distance;
if (dash_stream->keyframe_average_distance) {
/* Under-estimate the keyframe distance */
if (dash_stream->keyframe_average_distance > current_keyframe_distance)
dash_stream->keyframe_average_distance =
(dash_stream->keyframe_average_distance * 3 +
current_keyframe_distance) / 4;
else
dash_stream->keyframe_average_distance =
(dash_stream->keyframe_average_distance +
current_keyframe_distance * 3) / 4;
} else {
dash_stream->keyframe_average_distance = current_keyframe_distance;
}
GST_DEBUG_OBJECT (stream->pad,
"average keyframe sample size: %" G_GUINT64_FORMAT,
dash_stream->keyframe_average_size);
GST_DEBUG_OBJECT (stream->pad,
"average keyframe distance: %" GST_TIME_FORMAT " (%" GST_TIME_FORMAT
")", GST_TIME_ARGS (dash_stream->keyframe_average_distance),
GST_TIME_ARGS (current_keyframe_distance));
GST_DEBUG_OBJECT (stream->pad, "first sync sample after moof: %d",
dash_stream->first_sync_sample_after_moof);
}
return TRUE;
}
static GstFlowReturn
gst_dash_demux_handle_isobmff (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream)
{
GstDashDemuxStream *dash_stream = (GstDashDemuxStream *) stream;
GstFlowReturn ret = GST_FLOW_OK;
GstBuffer *buffer;
gboolean sidx_advance = FALSE;
/* We parse all ISOBMFF boxes of a (sub)fragment until the mdat. This covers
* at least moov, moof and sidx boxes. Once mdat is received we just output
* everything until the next (sub)fragment */
if (dash_stream->isobmff_parser.current_fourcc != GST_ISOFF_FOURCC_MDAT) {
gboolean sidx_seek_needed = FALSE;
ret = gst_dash_demux_parse_isobmff (demux, dash_stream, &sidx_seek_needed);
if (ret != GST_FLOW_OK)
return ret;
/* Go to selected segment if needed here */
if (sidx_seek_needed && !stream->downloading_index)
return GST_ADAPTIVE_DEMUX_FLOW_END_OF_FRAGMENT;
/* No mdat yet, let's get called again with the next boxes */
if (dash_stream->isobmff_parser.current_fourcc != GST_ISOFF_FOURCC_MDAT)
return ret;
/* Here we end up only if we're right at the mdat start */
/* Jump to the next sync sample. As we're doing chunked downloading
* here, just drop data until our chunk is over so we can reuse the
* HTTP connection instead of having to create a new one or
* reuse the data if the sync sample follows the moof */
if (dash_stream->active_stream->mimeType == GST_STREAM_VIDEO
&& gst_dash_demux_find_sync_samples (demux, stream) &&
GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (stream->demux)) {
guint idx = -1;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (GST_CLOCK_TIME_IS_VALID (dash_stream->target_time)) {
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
idx =
(dash_stream->target_time -
dash_stream->current_fragment_timestamp) /
dash_stream->current_fragment_keyframe_distance;
} else if (stream->segment.rate > 0) {
idx = 0;
}
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_DEBUG_OBJECT (stream->pad, "target %" GST_TIME_FORMAT " idx %d",
GST_TIME_ARGS (dash_stream->target_time), idx);
/* Figure out target time */
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
if (dash_stream->first_sync_sample_after_moof && idx == 0) {
/* If we're here, don't throw away data but collect sync
* sample while we're at it below. We're doing chunked
* downloading so might need to adjust the next chunk size for
* the remainder */
dash_stream->current_sync_sample = 0;
dashdemux: Improve key-unit trick mode downloading When dealing with key-unit trick mode downloads, the goal is to provide the best "Quality of Experience". This is achieved by: 1) maximizing the number of frames displayed per second 2) avoiding "stalling" as much as possible (i.e. not downloading and decoding frames fast enough) This implementation achives this by: 1) Knowing very precisely the current keyframe being download (i.e more accurate than at the fragment level which might contain more than one keyfram). This is the new "actual_position" variable introduced by this commit 2) Knowing the position of downstream (provided by QoS and stored in the adaptivedemuxstream qos_earliest_time variable) 3) Knowing how long it takes to request and fully download a keyframe (the average_download_time variable) Taking those 3 variables into account, whenever a keyframe has been pushed downstream we calculate a "target time" (target_time variable) which is the ideal next keyframe time to request so that: 1) It will be requested/downloaded/demuxed/decoded in time to be displayed without being too late 2) It will not be too far ahead that it would cause too few frames per second to be displayed. How far ahead we will request is inversily proportional to how close the actual position (actual_position) is from the downstream position (qos_earliest_time). The more is buffered between the source and the sink, the "closer" the target time will be, and therefore the more frames per seconds will be displayed (up to the limit of keyframes_per_second * absolute_rate).
2017-01-12 14:54:37 +00:00
GST_DEBUG_OBJECT (stream->pad, "Using first keyframe after header");
}
}
if (gst_adapter_available (dash_stream->adapter) == 0)
return ret;
/* We have some data from the mdat available in the adapter, handle it
* below in the push code */
} else {
/* Somewhere in the middle of the mdat */
}
/* At mdat */
if (dash_stream->sidx_parser.status == GST_ISOFF_SIDX_PARSER_FINISHED) {
guint64 sidx_end_offset =
dash_stream->sidx_base_offset +
SIDX_CURRENT_ENTRY (dash_stream)->offset +
SIDX_CURRENT_ENTRY (dash_stream)->size;
gboolean has_next = gst_dash_demux_stream_has_next_subfragment (stream);
gsize available;
/* Need to handle everything in the adapter according to the parsed SIDX
* and advance subsegments accordingly */
available = gst_adapter_available (dash_stream->adapter);
if (dash_stream->current_offset + available < sidx_end_offset) {
buffer = gst_adapter_take_buffer (dash_stream->adapter, available);
} else {
if (!has_next && sidx_end_offset <= dash_stream->current_offset) {
/* Drain all bytes, since there might be trailing bytes at the end of subfragment */
buffer = gst_adapter_take_buffer (dash_stream->adapter, available);
} else {
if (sidx_end_offset <= dash_stream->current_offset) {
/* This means a corrupted stream or a bug: ignoring bugs, it
* should only happen if the SIDX index is corrupt */
GST_ERROR_OBJECT (stream->pad, "Invalid SIDX state");
gst_adapter_clear (dash_stream->adapter);
return GST_FLOW_ERROR;
} else {
buffer =
gst_adapter_take_buffer (dash_stream->adapter,
sidx_end_offset - dash_stream->current_offset);
sidx_advance = TRUE;
}
}
}
} else {
/* Take it all and handle it further below */
buffer =
gst_adapter_take_buffer (dash_stream->adapter,
gst_adapter_available (dash_stream->adapter));
/* Attention: All code paths below need to update dash_stream->current_offset */
}
/* We're actually running in key-units trick mode */
if (dash_stream->active_stream->mimeType == GST_STREAM_VIDEO
&& dash_stream->moof_sync_samples
&& GST_ADAPTIVE_DEMUX_IN_TRICKMODE_KEY_UNITS (stream->demux)) {
if (dash_stream->current_sync_sample == -1) {
/* We're doing chunked downloading and wait for finishing the current
* chunk so we can jump to the first keyframe */
dash_stream->current_offset += gst_buffer_get_size (buffer);
gst_buffer_unref (buffer);
return GST_FLOW_OK;
} else {
GstDashStreamSyncSample *sync_sample =
&g_array_index (dash_stream->moof_sync_samples,
GstDashStreamSyncSample, dash_stream->current_sync_sample);
guint64 end_offset =
dash_stream->current_offset + gst_buffer_get_size (buffer);
/* Make sure to not download too much, this should only happen for
* the very first keyframe if it follows the moof */
if (dash_stream->current_offset >= sync_sample->end_offset + 1) {
dash_stream->current_offset += gst_buffer_get_size (buffer);
gst_buffer_unref (buffer);
return GST_FLOW_OK;
} else if (end_offset > sync_sample->end_offset + 1) {
guint64 remaining =
sync_sample->end_offset + 1 - dash_stream->current_offset;
GstBuffer *sub = gst_buffer_copy_region (buffer, GST_BUFFER_COPY_ALL, 0,
remaining);
gst_buffer_unref (buffer);
buffer = sub;
}
}
}
GST_BUFFER_OFFSET (buffer) = dash_stream->current_offset;
dash_stream->current_offset += gst_buffer_get_size (buffer);
GST_BUFFER_OFFSET_END (buffer) = dash_stream->current_offset;
ret = gst_adaptive_demux_stream_push_buffer (stream, buffer);
if (ret != GST_FLOW_OK)
return ret;
if (sidx_advance) {
ret =
gst_adaptive_demux_stream_advance_fragment (demux, stream,
SIDX_CURRENT_ENTRY (dash_stream)->duration);
if (ret != GST_FLOW_OK)
return ret;
/* If we still have data available, recurse and use it up if possible */
if (gst_adapter_available (dash_stream->adapter) > 0)
return gst_dash_demux_handle_isobmff (demux, stream);
}
return ret;
}
static GstFlowReturn
gst_dash_demux_data_received (GstAdaptiveDemux * demux,
GstAdaptiveDemuxStream * stream, GstBuffer * buffer)
{
GstDashDemuxStream *dash_stream = (GstDashDemuxStream *) stream;
GstFlowReturn ret = GST_FLOW_OK;
guint index_header_or_data;
if (stream->downloading_index)
index_header_or_data = 1;
else if (stream->downloading_header)
index_header_or_data = 2;
else
index_header_or_data = 3;
if (dash_stream->current_index_header_or_data != index_header_or_data) {
/* Clear pending data */
if (gst_adapter_available (dash_stream->adapter) != 0)
GST_ERROR_OBJECT (stream->pad,
"Had pending SIDX data after switch between index/header/data");
gst_adapter_clear (dash_stream->adapter);
dash_stream->current_index_header_or_data = index_header_or_data;
dash_stream->current_offset = -1;
}
if (dash_stream->current_offset == -1)
dash_stream->current_offset =
GST_BUFFER_OFFSET_IS_VALID (buffer) ? GST_BUFFER_OFFSET (buffer) : 0;
gst_adapter_push (dash_stream->adapter, buffer);
buffer = NULL;
if (dash_stream->is_isobmff || stream->downloading_index) {
/* SIDX index is also ISOBMMF */
ret = gst_dash_demux_handle_isobmff (demux, stream);
} else if (dash_stream->sidx_parser.status == GST_ISOFF_SIDX_PARSER_FINISHED) {
gsize available;
/* Not ISOBMFF but had a SIDX index. Does this even exist or work? */
while (ret == GST_FLOW_OK
&& ((available = gst_adapter_available (dash_stream->adapter)) > 0)) {
gboolean advance = FALSE;
guint64 sidx_end_offset =
dash_stream->sidx_base_offset +
SIDX_CURRENT_ENTRY (dash_stream)->offset +
SIDX_CURRENT_ENTRY (dash_stream)->size;
gboolean has_next = gst_dash_demux_stream_has_next_subfragment (stream);
if (dash_stream->current_offset + available < sidx_end_offset) {
buffer = gst_adapter_take_buffer (dash_stream->adapter, available);
} else {
if (!has_next && sidx_end_offset <= dash_stream->current_offset) {
/* Drain all bytes, since there might be trailing bytes at the end of subfragment */
buffer = gst_adapter_take_buffer (dash_stream->adapter, available);
} else {
if (sidx_end_offset <= dash_stream->current_offset) {
/* This means a corrupted stream or a bug: ignoring bugs, it
* should only happen if the SIDX index is corrupt */
GST_ERROR_OBJECT (stream->pad, "Invalid SIDX state");
gst_adapter_clear (dash_stream->adapter);
ret = GST_FLOW_ERROR;
break;
} else {
buffer =
gst_adapter_take_buffer (dash_stream->adapter,
sidx_end_offset - dash_stream->current_offset);
advance = TRUE;
}
}
}
GST_BUFFER_OFFSET (buffer) = dash_stream->current_offset;
GST_BUFFER_OFFSET_END (buffer) =
GST_BUFFER_OFFSET (buffer) + gst_buffer_get_size (buffer);
dash_stream->current_offset = GST_BUFFER_OFFSET_END (buffer);
ret = gst_adaptive_demux_stream_push_buffer (stream, buffer);
if (advance) {
if (has_next) {
GstFlowReturn new_ret;
new_ret =
gst_adaptive_demux_stream_advance_fragment (demux, stream,
SIDX_CURRENT_ENTRY (dash_stream)->duration);
/* only overwrite if it was OK before */
if (ret == GST_FLOW_OK)
ret = new_ret;
} else {
break;
}
}
}
} else {
/* this should be the main header, just push it all */
buffer = gst_adapter_take_buffer (dash_stream->adapter,
gst_adapter_available (dash_stream->adapter));
GST_BUFFER_OFFSET (buffer) = dash_stream->current_offset;
GST_BUFFER_OFFSET_END (buffer) =
GST_BUFFER_OFFSET (buffer) + gst_buffer_get_size (buffer);
dash_stream->current_offset = GST_BUFFER_OFFSET_END (buffer);
ret = gst_adaptive_demux_stream_push_buffer (stream, buffer);
}
return ret;
}
static void
gst_dash_demux_stream_free (GstAdaptiveDemuxStream * stream)
{
GstDashDemuxStream *dash_stream = (GstDashDemuxStream *) stream;
gst_isoff_sidx_parser_clear (&dash_stream->sidx_parser);
if (dash_stream->adapter)
g_object_unref (dash_stream->adapter);
if (dash_stream->moof)
gst_isoff_moof_box_free (dash_stream->moof);
if (dash_stream->moof_sync_samples)
g_array_free (dash_stream->moof_sync_samples, TRUE);
g_free (dash_stream->last_representation_id);
}
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
static GstDashDemuxClockDrift *
gst_dash_demux_clock_drift_new (GstDashDemux * demux)
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
{
GstDashDemuxClockDrift *clock_drift;
clock_drift = g_slice_new0 (GstDashDemuxClockDrift);
g_mutex_init (&clock_drift->clock_lock);
clock_drift->next_update =
GST_TIME_AS_USECONDS (gst_adaptive_demux_get_monotonic_time
(GST_ADAPTIVE_DEMUX_CAST (demux)));
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
return clock_drift;
}
static void
gst_dash_demux_clock_drift_free (GstDashDemuxClockDrift * clock_drift)
{
if (clock_drift) {
g_mutex_lock (&clock_drift->clock_lock);
if (clock_drift->ntp_clock)
g_object_unref (clock_drift->ntp_clock);
g_mutex_unlock (&clock_drift->clock_lock);
g_mutex_clear (&clock_drift->clock_lock);
g_slice_free (GstDashDemuxClockDrift, clock_drift);
}
}
/*
* The value attribute of the UTCTiming element contains a white-space
* separated list of servers that are recommended to be used in
* combination with the NTP protocol as defined in IETF RFC 5905 for
* getting the appropriate time.
*
* The DASH standard does not specify which version of NTP. This
* function only works with NTPv4 servers.
*/
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
static GstDateTime *
gst_dash_demux_poll_ntp_server (GstDashDemuxClockDrift * clock_drift,
gchar ** urls)
{
GstClockTime ntp_clock_time;
GDateTime *dt, *dt2;
if (!clock_drift->ntp_clock) {
GResolver *resolver;
GList *inet_addrs;
GError *err = NULL;
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
gchar *ip_addr;
resolver = g_resolver_get_default ();
/* We don't round-robin NTP servers. If the manifest specifies multiple
NTP time servers, select one at random */
clock_drift->selected_url = g_random_int_range (0, g_strv_length (urls));
GST_DEBUG ("Connecting to NTP time server %s",
urls[clock_drift->selected_url]);
inet_addrs = g_resolver_lookup_by_name (resolver,
urls[clock_drift->selected_url], NULL, &err);
g_object_unref (resolver);
if (!inet_addrs || g_list_length (inet_addrs) == 0) {
GST_ERROR ("Failed to resolve hostname of NTP server: %s",
err ? (err->message) : "unknown error");
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
if (inet_addrs)
g_resolver_free_addresses (inet_addrs);
if (err)
g_error_free (err);
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
return NULL;
}
ip_addr =
g_inet_address_to_string ((GInetAddress
*) (g_list_first (inet_addrs)->data));
clock_drift->ntp_clock = gst_ntp_clock_new ("dashntp", ip_addr, 123, 0);
g_free (ip_addr);
g_resolver_free_addresses (inet_addrs);
if (!clock_drift->ntp_clock) {
GST_ERROR ("Failed to create NTP clock");
return NULL;
}
if (!gst_clock_wait_for_sync (clock_drift->ntp_clock, 5 * GST_SECOND)) {
g_object_unref (clock_drift->ntp_clock);
clock_drift->ntp_clock = NULL;
GST_ERROR ("Failed to lock to NTP clock");
return NULL;
}
}
ntp_clock_time = gst_clock_get_time (clock_drift->ntp_clock);
if (ntp_clock_time == GST_CLOCK_TIME_NONE) {
GST_ERROR ("Failed to get time from NTP clock");
return NULL;
}
ntp_clock_time -= NTP_TO_UNIX_EPOCH * GST_SECOND;
dt = g_date_time_new_from_unix_utc (ntp_clock_time / GST_SECOND);
if (!dt) {
GST_ERROR ("Failed to create GstDateTime");
return NULL;
}
ntp_clock_time =
gst_util_uint64_scale (ntp_clock_time % GST_SECOND, 1000000, GST_SECOND);
dt2 = g_date_time_add (dt, ntp_clock_time);
g_date_time_unref (dt);
return gst_date_time_new_from_g_date_time (dt2);
}
struct Rfc5322TimeZone
{
const gchar *name;
gfloat tzoffset;
};
/*
Parse an RFC5322 (section 3.3) date-time from the Date: field in the
HTTP response.
See https://tools.ietf.org/html/rfc5322#section-3.3
*/
static GstDateTime *
gst_dash_demux_parse_http_head (GstDashDemuxClockDrift * clock_drift,
GstFragment * download)
{
static const gchar *months[] = { NULL, "Jan", "Feb", "Mar", "Apr",
"May", "Jun", "Jul", "Aug",
"Sep", "Oct", "Nov", "Dec", NULL
};
static const struct Rfc5322TimeZone timezones[] = {
{"Z", 0},
{"UT", 0},
{"GMT", 0},
{"BST", 1},
{"EST", -5},
{"EDT", -4},
{"CST", -6},
{"CDT", -5},
{"MST", -7},
{"MDT", -6},
{"PST", -8},
{"PDT", -7},
{NULL, 0}
};
GstDateTime *value = NULL;
const GstStructure *response_headers;
const gchar *http_date;
const GValue *val;
gint ret;
const gchar *pos;
gint year = -1, month = -1, day = -1, hour = -1, minute = -1, second = -1;
gchar zone[6];
gchar monthstr[4];
gfloat tzoffset = 0;
gboolean parsed_tz = FALSE;
g_return_val_if_fail (download != NULL, NULL);
g_return_val_if_fail (download->headers != NULL, NULL);
val = gst_structure_get_value (download->headers, "response-headers");
if (!val) {
return NULL;
}
response_headers = gst_value_get_structure (val);
http_date = gst_structure_get_string (response_headers, "Date");
if (!http_date) {
return NULL;
}
/* skip optional text version of day of the week */
pos = strchr (http_date, ',');
if (pos)
pos++;
else
pos = http_date;
ret =
sscanf (pos, "%02d %3s %04d %02d:%02d:%02d %5s", &day, monthstr, &year,
&hour, &minute, &second, zone);
if (ret == 7) {
gchar *z = zone;
gint i;
for (i = 1; months[i]; ++i) {
if (g_ascii_strncasecmp (months[i], monthstr, strlen (months[i])) == 0) {
month = i;
break;
}
}
for (i = 0; timezones[i].name && !parsed_tz; ++i) {
if (g_ascii_strncasecmp (timezones[i].name, z,
strlen (timezones[i].name)) == 0) {
tzoffset = timezones[i].tzoffset;
parsed_tz = TRUE;
}
}
if (!parsed_tz) {
gint hh, mm;
gboolean neg = FALSE;
/* check if it is in the form +-HHMM */
if (*z == '+' || *z == '-') {
if (*z == '+')
++z;
else if (*z == '-') {
++z;
neg = TRUE;
}
ret = sscanf (z, "%02d%02d", &hh, &mm);
if (ret == 2) {
tzoffset = hh;
tzoffset += mm / 60.0;
if (neg)
tzoffset = -tzoffset;
parsed_tz = TRUE;
}
}
}
/* Accept year in both 2 digit or 4 digit format */
if (year < 100)
year += 2000;
}
if (month > 0 && parsed_tz) {
value = gst_date_time_new (tzoffset,
year, month, day, hour, minute, second);
}
return value;
}
/*
The timing information is contained in the message body of the HTTP
response and contains a time value formatted according to NTP timestamp
format in IETF RFC 5905.
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Seconds |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Fraction |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
NTP Timestamp Format
*/
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
static GstDateTime *
gst_dash_demux_parse_http_ntp (GstDashDemuxClockDrift * clock_drift,
GstBuffer * buffer)
{
gint64 seconds;
guint64 fraction;
GDateTime *dt, *dt2;
GstMapInfo mapinfo;
/* See https://tools.ietf.org/html/rfc5905#page-12 for details of
the NTP Timestamp Format */
gst_buffer_map (buffer, &mapinfo, GST_MAP_READ);
if (mapinfo.size != 8) {
gst_buffer_unmap (buffer, &mapinfo);
return NULL;
}
seconds = GST_READ_UINT32_BE (mapinfo.data);
fraction = GST_READ_UINT32_BE (mapinfo.data + 4);
gst_buffer_unmap (buffer, &mapinfo);
fraction = gst_util_uint64_scale (fraction, 1000000,
G_GUINT64_CONSTANT (1) << 32);
/* subtract constant to convert from 1900 based time to 1970 based time */
seconds -= NTP_TO_UNIX_EPOCH;
dt = g_date_time_new_from_unix_utc (seconds);
dt2 = g_date_time_add (dt, fraction);
g_date_time_unref (dt);
return gst_date_time_new_from_g_date_time (dt2);
}
/*
The timing information is contained in the message body of the
HTTP response and contains a time value formatted according to
xs:dateTime as defined in W3C XML Schema Part 2: Datatypes specification.
*/
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
static GstDateTime *
gst_dash_demux_parse_http_xsdate (GstDashDemuxClockDrift * clock_drift,
GstBuffer * buffer)
{
GstDateTime *value = NULL;
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
GstMapInfo mapinfo;
/* the string from the server might not be zero terminated */
if (gst_buffer_map (buffer, &mapinfo, GST_MAP_READ)) {
gchar *str;
str = g_strndup ((const gchar *) mapinfo.data, mapinfo.size);
gst_buffer_unmap (buffer, &mapinfo);
value = gst_date_time_new_from_iso8601_string (str);
g_free (str);
}
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
return value;
}
static gboolean
gst_dash_demux_poll_clock_drift (GstDashDemux * demux)
{
GstDashDemuxClockDrift *clock_drift;
GDateTime *start = NULL, *end;
GstBuffer *buffer = NULL;
GstDateTime *value = NULL;
gboolean ret = FALSE;
gint64 now;
GstMPDUTCTimingType method;
gchar **urls;
g_return_val_if_fail (demux != NULL, FALSE);
g_return_val_if_fail (demux->clock_drift != NULL, FALSE);
clock_drift = demux->clock_drift;
now =
GST_TIME_AS_USECONDS (gst_adaptive_demux_get_monotonic_time
(GST_ADAPTIVE_DEMUX_CAST (demux)));
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
if (now < clock_drift->next_update) {
/*TODO: If a fragment fails to download in adaptivedemux, it waits
for a manifest reload before another attempt to fetch a fragment.
Section 10.8.6 of the DVB-DASH standard states that the DASH client
shall refresh the manifest and resynchronise to one of the time sources.
Currently the fact that the manifest refresh follows a download failure
does not make it into dashdemux. */
return TRUE;
}
urls = gst_mpd_client_get_utc_timing_sources (demux->client,
SUPPORTED_CLOCK_FORMATS, &method);
if (!urls) {
return FALSE;
}
/* Update selected_url just in case the number of URLs in the UTCTiming
element has shrunk since the last poll */
clock_drift->selected_url = clock_drift->selected_url % g_strv_length (urls);
g_mutex_lock (&clock_drift->clock_lock);
if (method == GST_MPD_UTCTIMING_TYPE_NTP) {
value = gst_dash_demux_poll_ntp_server (clock_drift, urls);
if (!value) {
GST_ERROR_OBJECT (demux, "Failed to fetch time from NTP server %s",
urls[clock_drift->selected_url]);
g_mutex_unlock (&clock_drift->clock_lock);
goto quit;
}
}
start =
gst_adaptive_demux_get_client_now_utc (GST_ADAPTIVE_DEMUX_CAST (demux));
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
if (!value) {
GstFragment *download;
gint64 range_start = 0, range_end = -1;
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
GST_DEBUG_OBJECT (demux, "Fetching current time from %s",
urls[clock_drift->selected_url]);
if (method == GST_MPD_UTCTIMING_TYPE_HTTP_HEAD) {
range_start = -1;
}
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
download =
gst_uri_downloader_fetch_uri_with_range (GST_ADAPTIVE_DEMUX_CAST
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
(demux)->downloader, urls[clock_drift->selected_url], NULL, TRUE, TRUE,
TRUE, range_start, range_end, NULL);
if (download) {
if (method == GST_MPD_UTCTIMING_TYPE_HTTP_HEAD && download->headers) {
value = gst_dash_demux_parse_http_head (clock_drift, download);
} else {
buffer = gst_fragment_get_buffer (download);
}
g_object_unref (download);
}
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
}
g_mutex_unlock (&clock_drift->clock_lock);
if (!value && !buffer) {
GST_ERROR_OBJECT (demux, "Failed to fetch time from %s",
urls[clock_drift->selected_url]);
goto quit;
}
end = gst_adaptive_demux_get_client_now_utc (GST_ADAPTIVE_DEMUX_CAST (demux));
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
if (!value && method == GST_MPD_UTCTIMING_TYPE_HTTP_NTP) {
value = gst_dash_demux_parse_http_ntp (clock_drift, buffer);
} else if (!value) {
/* GST_MPD_UTCTIMING_TYPE_HTTP_XSDATE or GST_MPD_UTCTIMING_TYPE_HTTP_ISO */
value = gst_dash_demux_parse_http_xsdate (clock_drift, buffer);
}
if (buffer)
gst_buffer_unref (buffer);
if (value) {
GTimeSpan download_duration = g_date_time_difference (end, start);
GDateTime *client_now, *server_now;
/* We don't know when the server sampled its clock, but we know
it must have been before "end" and probably after "start".
A reasonable estimate is to use (start+end)/2
*/
client_now = g_date_time_add (start, download_duration / 2);
server_now = gst_date_time_to_g_date_time (value);
/* If gst_date_time_new_from_iso8601_string is given an unsupported
ISO 8601 format, it can return a GstDateTime that is not valid,
which causes gst_date_time_to_g_date_time to return NULL */
if (server_now) {
g_mutex_lock (&clock_drift->clock_lock);
clock_drift->clock_compensation =
g_date_time_difference (server_now, client_now);
g_mutex_unlock (&clock_drift->clock_lock);
GST_DEBUG_OBJECT (demux,
"Difference between client and server clocks is %lfs",
((double) clock_drift->clock_compensation) / 1000000.0);
g_date_time_unref (server_now);
ret = TRUE;
} else {
GST_ERROR_OBJECT (demux, "Failed to parse DateTime from server");
}
g_date_time_unref (client_now);
gst_date_time_unref (value);
} else {
GST_ERROR_OBJECT (demux, "Failed to parse DateTime from server");
}
g_date_time_unref (end);
quit:
if (start)
g_date_time_unref (start);
/* if multiple URLs were specified, use a simple round-robin to
poll each server */
g_mutex_lock (&clock_drift->clock_lock);
if (method == GST_MPD_UTCTIMING_TYPE_NTP) {
clock_drift->next_update = now + FAST_CLOCK_UPDATE_INTERVAL;
} else {
clock_drift->selected_url =
(1 + clock_drift->selected_url) % g_strv_length (urls);
if (ret) {
clock_drift->next_update = now + SLOW_CLOCK_UPDATE_INTERVAL;
} else {
clock_drift->next_update = now + FAST_CLOCK_UPDATE_INTERVAL;
}
}
g_mutex_unlock (&clock_drift->clock_lock);
return ret;
}
static GTimeSpan
gst_dash_demux_get_clock_compensation (GstDashDemux * demux)
{
GTimeSpan rv = 0;
if (demux->clock_drift) {
g_mutex_lock (&demux->clock_drift->clock_lock);
rv = demux->clock_drift->clock_compensation;
g_mutex_unlock (&demux->clock_drift->clock_lock);
}
GST_LOG_OBJECT (demux, "Clock drift %" GST_STIME_FORMAT, GST_STIME_ARGS (rv));
return rv;
}
static GDateTime *
gst_dash_demux_get_server_now_utc (GstDashDemux * demux)
{
GDateTime *client_now;
GDateTime *server_now;
client_now =
gst_adaptive_demux_get_client_now_utc (GST_ADAPTIVE_DEMUX_CAST (demux));
server_now =
g_date_time_add (client_now,
dashdemux: add support for UTCTiming elements for clock drift compensation Unless the DASH client can compensate for the difference between its clock and the clock used by the server, the client might request fragments that either not yet on the server or fragments that have already been expired from the server. This is an issue because these requests can propagate all the way back to the origin ISO/IEC 23009-1:2014/Amd 1 [PDAM1] defines a new UTCTiming element to allow a DASH client to track the clock used by the server generating the DASH stream. Multiple UTCTiming elements might be present, to indicate support for multiple methods of UTC time gathering. Each element can contain a white space separated list of URLs that can be contacted to discover the UTC time from the server's perspective. This commit provides parsing of UTCTiming elements, unit tests of this parsing and a function to poll a time server. This function supports the following methods: urn:mpeg:dash:utc:ntp:2014 urn:mpeg:dash:utc:http-xsdate:2014 urn:mpeg:dash:utc:http-iso:2014 urn:mpeg:dash:utc:http-ntp:2014 The manifest update task is used to poll the clock time server, to save having to create a new thread. When choosing the starting fragment number and when waiting for a fragment to become available, the difference between the server's idea of UTC and the client's idea of UTC is taken into account. For example, if the server's time is behind the client's idea of UTC, we wait for longer before requesting a fragment [PDAM1]: http://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=66068 dashdemux: support NTP time servers in UTCTiming elements Use the gst_ntp_clock to support the use of an NTP server. https://bugzilla.gnome.org/show_bug.cgi?id=752413
2015-07-15 10:56:13 +00:00
gst_dash_demux_get_clock_compensation (demux));
g_date_time_unref (client_now);
return server_now;
}