Nicolas Dufresne schreef op 2024-03-06 22:14:
<snip>
> > + /* vb2 timestamps in nsecs, mmal in usecs */
> > + vb2->vb2_buf.timestamp = mmal_buf->pts * 1000;
>
> M2M drivers are not allowed to "interpret" the timesamp field. They can
> only
> copy it over. This scaling of the timestamp seems problematic and may
> possibly
> impair some userspace ability to match in and out buffers. Is there
> some other
> user data that could be use to maintain the lower bits ?
>
> I realized that it should not have worked for GStreamer, since it uses
> a frame
> counter as timestamp (and no real timestamp), but accidentally, it only
> uses
> tv_sec member (with a implicit cast from signed to unsigned, but lets
> not grab
> other projects issues in, unlike stateless, there is a single
> representation of
> it, so its less of an issue).
>
> https://gitlab.freedesktop.org/gstreamer/gstreamer/-/blob/main/subprojects/gst-plugins-good/sys/v4l2/gstv4l2bufferpool.c?ref_type=heads#L1179
>
> There is no requirement in the spec that tells userspace using forward
> counters
> to only use the seconds, so I suppose this is far from ideal.
You mean that vb2_buf.timestamp is not necessarily a time-based
timestamp? I'll have to see how mmal is using the pts.
Correct, its opaque data from user, the spec only defines the transfer
rules for
that data in the case that a frame is duplicated or merged.
So, I asked downstream to (if possible) handle mmal timestamps opaquely,
which would simplify this code a lot and fix your GStreamer issues:
https://github.com/raspberrypi/firmware/issues/1883 . We'll see if it's
possible or not; if not I'll have to look for a way to present a time to
mmal while keeping the opaque data intact.
Maarten
