Re: [PATCH] media: coda: Keep metas sync with hardware fifo

From: Philipp Zabel
Date: Fri Nov 12 2021 - 04:30:35 EST


Hi Benjamin,

On Fri, 2021-11-12 at 09:36 +0100, Benjamin Gaignard wrote:
> Le 11/11/2021 à 18:03, Philipp Zabel a écrit :
> > Hi Benjamin,
> >
> > On Mon, 2021-11-08 at 15:24 +0100, Benjamin Gaignard wrote:
> > > After updating the output fifo position be sure that metas are also
> > > synchronised with this position.
> > >
> > > Signed-off-by: Benjamin Gaignard <benjamin.gaignard@xxxxxxxxxxxxx>
> > > ---
> > >   drivers/media/platform/coda/coda-bit.c | 6 ++++++
> > >   1 file changed, 6 insertions(+)
> > >
> > > diff --git a/drivers/media/platform/coda/coda-bit.c b/drivers/media/platform/coda/coda-bit.c
> > > index c484c008ab02..28c56286b0de 100644
> > > --- a/drivers/media/platform/coda/coda-bit.c
> > > +++ b/drivers/media/platform/coda/coda-bit.c
> > > @@ -2315,6 +2315,12 @@ static void coda_finish_decode(struct coda_ctx *ctx)
> > >    /* Update kfifo out pointer from coda bitstream read pointer */
> > >    coda_kfifo_sync_from_device(ctx);
> > >
> > > + /*
> > > + * After updating the read pointer, we need to check if
> > > + * any metas are consumed and should be released.
> > > + */
> > > + coda_decoder_drop_used_metas(ctx);
> > > +
> > This doesn't look right. If you drop all metas seen by the decoder right
> > away, they can't be copied into the decoded picture's meta slot later in
> > this function. I'd expect you run into the "empty timestamp list!"
> > errors if you do this.
>
> Hi Philipp,
>
> I don't run into the "empty timestamp list!" errors.
> The only case I have seen metas been dropped here it is when
> an invalid/incomplete frame has been send into the decoder.
> When that happens I don't see any interrupt or error message but
> the erroneous frame stay in the list.
> If that occur 4 times (I'm using CODA 960) then the decoder hang.
> Dropping the metas at this moment solve this problem.
>
> Maybe you can guide me to a better solution ?

Hmm, the current code silently assumes that in the good case there's
exactly one used meta (it just picks the first one from the list if
decoded_idx is set).

We could list_cut_before() the used metas into a local list, let the
decoded_idx code pick the timestamp from that list, and then drop the
whole local list of used metas on return from coda_finish_decode().

regards
Philipp