I have one stream whose codec is xvid encoded, which have B picture inside.

I looked at the way of gstreamer native mkv demux to give time stamp:

buffer time = (cluster_time + time) * demux->time_scale;

So the gstreamer native mkv demux will not consider the display order of
frames. 

So in I P B B case (decoding order):
the P timestamp can be 2 second, and B time stamp can be 3 second, even though
the B should be displayed before the P.

And if downstream element think coming buffer time stamp is display time information, then it will be wrong.

So here the 1st question: whether the buffer time stamp from demux element is decoding time stamp or presentation time stamp?

I backed to ffmpeg to see how ffmpeg handle the divx in mkv container: I saw it will detect the frame type and only take B frame's decoding time (the same way as gstreamer native mkv demux caculate from cluster time and block time) as presentation time. For P and I frame, invalid time stamp is given.

Here the 2nd question: whether gstreamer native mkv demux should adopt the same strategy as ffmpeg? detect frame type and give correct PTS as buffer time stamp?