Solve the problem of inaccurate HLS TS duration. v5.0.187 (#3824)

Dockerfile

@@ -57,5 +57,5 @@ RUN ldd /usr/local/srs/objs/ffmpeg/bin/ffmpeg && \
 # Default workdir and command.
 WORKDIR /usr/local/srs
 ENV SRS_DAEMON=off
-CMD ["./objs/srs", "-c", "conf/srs.conf"]
+CMD ["./objs/srs", "-c", "conf/docker.conf"]
 

trunk/conf/full.conf

@@ -1747,27 +1747,6 @@ vhost hls.srs.com {
         # default: off
         enabled on;
 
-        # Whether enable hls_ctx for HLS streaming, for which we create a "fake" connection for HTTP API and callback.
-        # For each HLS streaming session, we use a child m3u8 with a session identified by query "hls_ctx", it simply
-        # work as the session id.
-        # Once the HLS streaming session is created, we will cleanup it when timeout in 2*hls_window seconds. So it
-        # takes a long time period to identify the timeout.
-        # Now we got a HLS stremaing session, just like RTMP/WebRTC/HTTP-FLV streaming, we're able to stat the session
-        # as a "fake" connection, do HTTP callback when start playing the HLS streaming. You're able to do querying and
-        # authentication.
-        # Note that it will make NGINX edge cache always missed, so never enable HLS streaming if use NGINX edges.
-        # Overwrite by env SRS_VHOST_HLS_HLS_CTX for all vhosts.
-        # Default: on
-        hls_ctx on;
-        # For HLS pseudo streaming, whether enable the session for each TS segment.
-        # If enabled, SRS HTTP API will show the statistics about HLS streaming bandwidth, both m3u8 and ts file. Please
-        # note that it also consumes resource, because each ts file should be served by SRS, all NGINX cache will be
-        # missed because we add session id to each ts file.
-        # Note that it will make NGINX edge cache always missed, so never enable HLS streaming if use NGINX edges.
-        # Overwrite by env SRS_VHOST_HLS_HLS_TS_CTX for all vhosts.
-        # Default: on
-        hls_ts_ctx on;
-
         # the hls fragment in seconds, the duration of a piece of ts.
         # Overwrite by env SRS_VHOST_HLS_HLS_FRAGMENT for all vhosts.
         # default: 10
@@ -1776,15 +1755,15 @@ vhost hls.srs.com {
         #   EXT-X-TARGETDURATION = hls_td_ratio * hls_fragment // init
         #   EXT-X-TARGETDURATION = max(ts_duration, EXT-X-TARGETDURATION) // for each ts
         # Overwrite by env SRS_VHOST_HLS_HLS_TD_RATIO for all vhosts.
-        # default: 1.5
+        # default: 1.0
-        hls_td_ratio 1.5;
+        hls_td_ratio 1.0;
         # the audio overflow ratio.
         # for pure audio, the duration to reap the segment.
-        # for example, the hls_fragment is 10s, hls_aof_ratio is 2.0,
+        # for example, the hls_fragment is 10s, hls_aof_ratio is 1.2,
-        # the segment will reap to 20s for pure audio.
+        # the segment will reap to 12s for pure audio.
         # Overwrite by env SRS_VHOST_HLS_HLS_AOF_RATIO for all vhosts.
-        # default: 2.0
+        # default: 1.2
-        hls_aof_ratio 2.0;
+        hls_aof_ratio 1.2;
         # the hls window in seconds, the number of ts in m3u8.
         # Overwrite by env SRS_VHOST_HLS_HLS_WINDOW for all vhosts.
         # default: 60
@@ -1833,13 +1812,6 @@ vhost hls.srs.com {
         # Overwrite by env SRS_VHOST_HLS_HLS_TS_FILE for all vhosts.
         # default: [app]/[stream]-[seq].ts
         hls_ts_file [app]/[stream]-[seq].ts;
-        # whether use floor for the hls_ts_file path generation.
-        # if on, use floor(timestamp/hls_fragment) as the variable [timestamp],
-        #       and use enhanced algorithm to calc deviation for segment.
-        # @remark when floor on, recommend the hls_segment>=2*gop.
-        # Overwrite by env SRS_VHOST_HLS_HLS_TS_FLOOR for all vhosts.
-        # default: off
-        hls_ts_floor off;
         # the hls entry prefix, which is base url of ts url.
         # for example, the prefix is:
         #         http://your-server/
@@ -1875,20 +1847,48 @@ vhost hls.srs.com {
         # @remark 0 to disable dispose for publisher.
         # @remark apply for publisher timeout only, while "etc/init.d/srs stop" always dispose hls.
         # Overwrite by env SRS_VHOST_HLS_HLS_DISPOSE for all vhosts.
-        # default: 0
+        # default: 120
-        hls_dispose 0;
+        hls_dispose 120;
-        # the max size to notify hls,
-        # to read max bytes from ts of specified cdn network,
-        # @remark only used when on_hls_notify is config.
-        # Overwrite by env SRS_VHOST_HLS_HLS_NB_NOTIFY for all vhosts.
-        # default: 64
-        hls_nb_notify 64;
         # whether wait keyframe to reap segment,
         # if off, reap segment when duration exceed the fragment,
         # if on, reap segment when duration exceed and got keyframe.
         # Overwrite by env SRS_VHOST_HLS_HLS_WAIT_KEYFRAME for all vhosts.
         # default: on
         hls_wait_keyframe on;
+        # whether use floor for the hls_ts_file path generation.
+        # if on, use floor(timestamp/hls_fragment) as the variable [timestamp],
+        #       and use enhanced algorithm to calc deviation for segment.
+        # @remark when floor on, recommend the hls_segment>=2*gop.
+        # Overwrite by env SRS_VHOST_HLS_HLS_TS_FLOOR for all vhosts.
+        # default: off
+        hls_ts_floor off;
+        # the max size to notify hls,
+        # to read max bytes from ts of specified cdn network,
+        # @remark only used when on_hls_notify is config.
+        # Overwrite by env SRS_VHOST_HLS_HLS_NB_NOTIFY for all vhosts.
+        # default: 64
+        hls_nb_notify 64;
+
+        # Whether enable hls_ctx for HLS streaming, for which we create a "fake" connection for HTTP API and callback.
+        # For each HLS streaming session, we use a child m3u8 with a session identified by query "hls_ctx", it simply
+        # work as the session id.
+        # Once the HLS streaming session is created, we will cleanup it when timeout in 2*hls_window seconds. So it
+        # takes a long time period to identify the timeout.
+        # Now we got a HLS stremaing session, just like RTMP/WebRTC/HTTP-FLV streaming, we're able to stat the session
+        # as a "fake" connection, do HTTP callback when start playing the HLS streaming. You're able to do querying and
+        # authentication.
+        # Note that it will make NGINX edge cache always missed, so never enable HLS streaming if use NGINX edges.
+        # Overwrite by env SRS_VHOST_HLS_HLS_CTX for all vhosts.
+        # Default: on
+        hls_ctx on;
+        # For HLS pseudo streaming, whether enable the session for each TS segment.
+        # If enabled, SRS HTTP API will show the statistics about HLS streaming bandwidth, both m3u8 and ts file. Please
+        # note that it also consumes resource, because each ts file should be served by SRS, all NGINX cache will be
+        # missed because we add session id to each ts file.
+        # Note that it will make NGINX edge cache always missed, so never enable HLS streaming if use NGINX edges.
+        # Overwrite by env SRS_VHOST_HLS_HLS_TS_CTX for all vhosts.
+        # Default: on
+        hls_ts_ctx on;
 
         # whether using AES encryption.
         # Overwrite by env SRS_VHOST_HLS_HLS_KEYS for all vhosts.

trunk/conf/hls.origin.conf

@@ -12,6 +12,7 @@ vhost __defaultVhost__ {
         enabled         on;
         # Note that it will make NGINX edge cache always missed, so never enable HLS streaming if use NGINX edges.
         hls_ctx off;
+        hls_ts_ctx off;
     }
 }
 

trunk/conf/hls.realtime.conf

@@ -14,8 +14,7 @@ http_server {
 vhost __defaultVhost__ {
     hls {
         enabled         on;
-        hls_fragment    0.2;
+        hls_fragment    2;
-        hls_window      2;
+        hls_window      10;
-        hls_wait_keyframe       off;
     }
 }

trunk/doc/CHANGELOG.md

@@ -7,6 +7,7 @@ The changelog for SRS.
 <a name="v5-changes"></a>
 
 ## SRS 5.0 Changelog
+* v5.0, 2023-10-08, Merge [#3824](https://github.com/ossrs/srs/pull/3824): Solve the problem of inaccurate HLS TS duration. v5.0.187 (#3824)
 * v5.0, 2023-10-08, Merge [#3815](https://github.com/ossrs/srs/pull/3815): Use new cache image name. v5.0.186 (#3815)
 * v5.0, 2023-09-28, Merge [#3816](https://github.com/ossrs/srs/pull/3816): cherry-pick from develop, for srt utest. v5.0.185 (#3816)
 * v5.0, 2023-09-21, Merge [#3806](https://github.com/ossrs/srs/pull/3806): Build: Support sys-ssl for srt. v5.0.184 (#3806)

trunk/src/app/srs_app_config.cpp

@@ -6973,7 +6973,7 @@ double SrsConfig::get_hls_td_ratio(string vhost)
 {
     SRS_OVERWRITE_BY_ENV_FLOAT("srs.vhost.hls.hls_td_ratio"); // SRS_VHOST_HLS_HLS_TD_RATIO
 
-    static double DEFAULT = 1.5;
+    static double DEFAULT = 1.0;
     
     SrsConfDirective* conf = get_hls(vhost);
     if (!conf) {
@@ -6992,7 +6992,7 @@ double SrsConfig::get_hls_aof_ratio(string vhost)
 {
     SRS_OVERWRITE_BY_ENV_FLOAT("srs.vhost.hls.hls_aof_ratio"); // SRS_VHOST_HLS_HLS_AOF_RATIO
 
-    static double DEFAULT = 2.0;
+    static double DEFAULT = 1.2;
     
     SrsConfDirective* conf = get_hls(vhost);
     if (!conf) {
@@ -7183,7 +7183,7 @@ srs_utime_t SrsConfig::get_hls_dispose(string vhost)
 {
     SRS_OVERWRITE_BY_ENV_SECONDS("srs.vhost.hls.hls_dispose"); // SRS_VHOST_HLS_HLS_DISPOSE
 
-    static srs_utime_t DEFAULT = 0;
+    static srs_utime_t DEFAULT = 120 * SRS_UTIME_SECONDS;
     
     SrsConfDirective* conf = get_hls(vhost);
     if (!conf) {

trunk/src/app/srs_app_hls.cpp

@@ -519,9 +519,12 @@ bool SrsHlsMuxer::is_segment_overflow()
         return false;
     }
     
-    // use N% deviation, to smoother.
+    // Use N% deviation, to smoother.
     srs_utime_t deviation = hls_ts_floor? SRS_HLS_FLOOR_REAP_PERCENT * deviation_ts * hls_fragment : 0;
-    return current->duration() >= hls_fragment + deviation;
+
+    // Keep in mind that we use max_td for the base duration, not the hls_fragment. To calculate
+    // max_td, multiply hls_fragment by hls_td_ratio.
+    return current->duration() >= max_td + deviation;
 }
 
 bool SrsHlsMuxer::wait_keyframe()
@@ -564,7 +567,7 @@ srs_error_t SrsHlsMuxer::flush_audio(SrsTsMessageCache* cache)
     }
     
     // update the duration of segment.
-    current->append(cache->audio->pts / 90);
+    update_duration(cache->audio->dts);
 
     if ((err = current->tscw->write_audio(cache->audio)) != srs_success) {
         return srs_error_wrap(err, "hls: write audio");
@@ -593,7 +596,7 @@ srs_error_t SrsHlsMuxer::flush_video(SrsTsMessageCache* cache)
     srs_assert(current);
     
     // update the duration of segment.
-    current->append(cache->video->dts / 90);
+    update_duration(cache->video->dts);
 
     if ((err = current->tscw->write_video(cache->video)) != srs_success) {
         return srs_error_wrap(err, "hls: write video");
@@ -605,6 +608,11 @@ srs_error_t SrsHlsMuxer::flush_video(SrsTsMessageCache* cache)
     return err;
 }
 
+void SrsHlsMuxer::update_duration(uint64_t dts)
+{
+    current->append(dts / 90);
+}
+
 srs_error_t SrsHlsMuxer::segment_close()
 {
     srs_error_t err = do_segment_close();
@@ -635,9 +643,9 @@ srs_error_t SrsHlsMuxer::do_segment_close()
     // valid, add to segments if segment duration is ok
     // when too small, it maybe not enough data to play.
     // when too large, it maybe timestamp corrupt.
-    // make the segment more acceptable, when in [min, max_td * 2], it's ok.
+    // make the segment more acceptable, when in [min, max_td * 3], it's ok.
     bool matchMinDuration = current->duration() >= SRS_HLS_SEGMENT_MIN_DURATION;
-    bool matchMaxDuration = current->duration() <= max_td * 2 * 1000;
+    bool matchMaxDuration = current->duration() <= max_td * 3 * 1000;
     if (matchMinDuration && matchMaxDuration) {
         // rename from tmp to real path
         if ((err = current->rename()) != srs_success) {
@@ -901,6 +909,7 @@ srs_error_t SrsHlsController::on_publish(SrsRequest* req)
     std::string app = req->app;
 
     srs_utime_t hls_fragment = _srs_config->get_hls_fragment(vhost);
+    double hls_td_ratio = _srs_config->get_hls_td_ratio(vhost);
     srs_utime_t hls_window = _srs_config->get_hls_window(vhost);
     
     // get the hls m3u8 ts list entry prefix config
@@ -944,9 +953,9 @@ srs_error_t SrsHlsController::on_publish(SrsRequest* req)
     // This config item is used in SrsHls, we just log its value here.
     bool hls_dts_directly = _srs_config->get_vhost_hls_dts_directly(req->vhost);
 
-    srs_trace("hls: win=%dms, frag=%dms, prefix=%s, path=%s, m3u8=%s, ts=%s, aof=%.2f, floor=%d, clean=%d, waitk=%d, dispose=%dms, dts_directly=%d",
+    srs_trace("hls: win=%dms, frag=%dms, prefix=%s, path=%s, m3u8=%s, ts=%s, tdr=%.2f, aof=%.2f, floor=%d, clean=%d, waitk=%d, dispose=%dms, dts_directly=%d",
         srsu2msi(hls_window), srsu2msi(hls_fragment), entry_prefix.c_str(), path.c_str(), m3u8_file.c_str(), ts_file.c_str(),
-        hls_aof_ratio, ts_floor, cleanup, wait_keyframe, srsu2msi(hls_dispose), hls_dts_directly);
+        hls_td_ratio, hls_aof_ratio, ts_floor, cleanup, wait_keyframe, srsu2msi(hls_dispose), hls_dts_directly);
     
     return err;
 }
@@ -997,6 +1006,10 @@ srs_error_t SrsHlsController::write_audio(SrsAudioFrame* frame, int64_t pts)
         return srs_error_wrap(err, "hls: cache audio");
     }
 
+    // First, update the duration of the segment, as we might reap the segment. The duration should
+    // cover from the first frame to the last frame.
+    muxer->update_duration(tsmc->audio->dts);
+    
     // reap when current source is pure audio.
     // it maybe changed when stream info changed,
     // for example, pure audio when start, audio/video when publishing,
@@ -1039,6 +1052,10 @@ srs_error_t SrsHlsController::write_video(SrsVideoFrame* frame, int64_t dts)
         return srs_error_wrap(err, "hls: cache video");
     }
 
+    // First, update the duration of the segment, as we might reap the segment. The duration should
+    // cover from the first frame to the last frame.
+    muxer->update_duration(tsmc->video->dts);
+    
     // when segment overflow, reap if possible.
     if (muxer->is_segment_overflow()) {
         // do reap ts if any of:

trunk/src/app/srs_app_hls.hpp

@@ -200,6 +200,10 @@ public:
     virtual bool pure_audio();
     virtual srs_error_t flush_audio(SrsTsMessageCache* cache);
     virtual srs_error_t flush_video(SrsTsMessageCache* cache);
+    // When flushing video or audio, we update the duration. But, we should also update the
+    // duration before closing the segment. Keep in mind that it's fine to update the duration
+    // several times using the same dts timestamp.
+    void update_duration(uint64_t dts);
     // Close segment(ts).
     virtual srs_error_t segment_close();
 private:

trunk/src/core/srs_core_version5.hpp

@@ -9,6 +9,6 @@
 
 #define VERSION_MAJOR       5
 #define VERSION_MINOR       0
-#define VERSION_REVISION    186
+#define VERSION_REVISION    187
 
 #endif

trunk/src/utest/srs_utest_config.cpp

@@ -2085,7 +2085,7 @@ VOID TEST(ConfigUnitTest, CheckDefaultValuesVhost)
 
     if (true) {
         HELPER_ASSERT_SUCCESS(conf.parse(_MIN_OK_CONF));
-        EXPECT_EQ(0, (int)conf.get_hls_dispose(""));
+        EXPECT_EQ(120 * SRS_UTIME_SECONDS, (int)conf.get_hls_dispose(""));
         EXPECT_EQ(10 * SRS_UTIME_SECONDS, conf.get_hls_fragment(""));
         EXPECT_EQ(60 * SRS_UTIME_SECONDS, conf.get_hls_window(""));