Wrong timestamp calculation causing RTCP clock skew (thanks Guido Fischer for the fix!)

[bennylp]

Quoting Guido's post on the list:

"There seems to be a bug in "rtcp.c". The rtt-time was faulty at some point and a few times, when enabling the trace in rtcp.c, I got "Internal RTCP NTP clock skew detected"-failures (time now < time of lsr).

In my understanding the problem is in pjmedia_rtcp_get_ntp_time:

    /* Calculate seconds fractions */
    ts.u64 %= sess->ts_freq.u64;
    pj_assert(ts.u64 < sess->ts_freq.u64);
    ts.u64 = (ts.u64 << 32) / sess->ts_freq.u64;

When calculating ntp->hi, this is done hooked on its base (ts.u64 - sess->ts_base.u64). I think this should also be done when calculating the fractions.

    /* Calculate seconds fractions */
    ts.u64 = (ts.u64 - sess->ts_base.u64) % sess->ts_freq.u64;
    pj_assert(ts.u64 < sess->ts_freq.u64);
    ts.u64 = (ts.u64 << 32) / sess->ts_freq.u64;

Worked fine for me. No more clock skews and noticeable rtt-time failures."

Kudos Guido Fischer for fixing this two and half year old bug!

[bennylp]

The corresponding ticket for 1.0.x branch is ticket #697

[bennylp]

Fixed in r2422