add explanatory comment about timestamp compensation (bug #4763)
authorKevin P. Fleming <kpfleming@digium.com>
Mon, 25 Jul 2005 23:13:20 +0000 (23:13 +0000)
committerKevin P. Fleming <kpfleming@digium.com>
Mon, 25 Jul 2005 23:13:20 +0000 (23:13 +0000)
git-svn-id: https://origsvn.digium.com/svn/asterisk/trunk@6217 65c4cc65-6c06-0410-ace0-fbb531ad65f3

channels/chan_iax2.c

index 93d2726..cdb640e 100755 (executable)
@@ -3420,11 +3420,23 @@ static unsigned int calc_timestamp(struct chan_iax2_pvt *p, unsigned int ts, str
                if (voice) {
                        /* On a voice frame, use predicted values if appropriate */
                        if (p->notsilenttx && abs(ms - p->nextpred) <= MAX_TIMESTAMP_SKEW) {
-                               /* Adjust our txcore, keeping voice and 
-                                       non-voice synchronized */
-                               /* We need someone who understands this code to comment here on
-                                  why the 'adjust' value is handled as if it was in units
-                                  of 10,000 microseconds, instead of milliseconds
+                               /* Adjust our txcore, keeping voice and non-voice synchronized */
+                               /* AN EXPLANATION:
+                                  When we send voice, we usually send "calculated" timestamps worked out
+                                  on the basis of the number of samples sent. When we send other frames,
+                                  we usually send timestamps worked out from the real clock.
+                                  The problem is that they can tend to drift out of step because the 
+                                  source channel's clock and our clock may not be exactly at the same rate.
+                                  We fix this by continuously "tweaking" p->offset.  p->offset is "time zero"
+                                  for this call.  Moving it adjusts timestamps for non-voice frames.
+                                  We make the adjustment in the style of a moving average.  Each time we
+                                  adjust p->offset by 10% of the difference between our clock-derived
+                                  timestamp and the predicted timestamp.  That's why you see "10000"
+                                  below even though IAX2 timestamps are in milliseconds.
+                                  The use of a moving average avoids offset moving too radically.
+                                  Generally, "adjust" roams back and forth around 0, with offset hardly
+                                  changing at all.  But if a consistent different starts to develop it
+                                  will be eliminated over the course of 10 frames (200-300msecs) 
                                */
                                adjust = (ms - p->nextpred);
                                if (adjust < 0)