Re: [PATCH v2 0/3] Add Amlogic video decoder driver

[Maxime Jourdan <mjourdan@baylibre.com>]

Le ven. 21 sept. 2018 à 12:51, Hans Verkuil <hverkuil@xs4all.nl> a écrit :
>
> On 09/17/18 18:36, Maxime Jourdan wrote:
> > 2018-09-17 16:51 GMT+02:00 Hans Verkuil <hverkuil@xs4all.nl>:
> >> On 09/11/2018 05:09 PM, Maxime Jourdan wrote:
> >>>  - Moved the single instance check (returning -EBUSY) to start/stop streaming
> >>>  The check was previously in queue_setup but there was no great location to
> >>>  clear it except for .close().
> >>
> >> Actually, you can clear it by called VIDIOC_REQBUFS with count set to 0. That
> >> freed all buffers and clears this.
> >>
> >> Now, the difference between queue_setup and start/stop streaming is that if you
> >> do this in queue_setup you'll know early on that the device is busy. It is
> >> reasonable to assume that you only allocate buffers when you also want to start
> >> streaming, so that it a good place to know this quickly.
> >>
> >> Whereas with start_streaming you won't know until you call STREAMON, or even later
> >> if you start streaming with no buffers queued, since start_streaming won't
> >> be called until you have at least 'min_buffers_needed' buffers queued (1 for this
> >> driver). So in that case EBUSY won't be returned until the first VIDIOC_QBUF.
> >>
> >> My preference is to check this in queue_setup, but it is up to you to decide.
> >> Just be aware of the difference between the two options.
> >>
> >> Regards,
> >>
> >>         Hans
> >
> > I could for instance keep track of which queue(s) have been called
> > with queue_setup, catch calls to VIDIOC_REQBUFS with count set to 0,
> > and clear the current session once both queues have been reset ?
>
> I see your point, this is rather awkward. The real problem here is that
> we don't have a 'queue_free' callback. If we'd had that this would be
> a lot easier.
>
> In any case, I am dropping my objections to doing this in start/stop_streaming.

Ack.

> > You leverage another issue with min_buffers_needed. It's indeed set to
> > 1 but this value is wrong for the CAPTURE queue. The problem is that
> > this value changes depending on the codec and the amount of CAPTURE
> > buffers requested by userspace.
> > Ultimately I want it set to the total amount of CAPTURE buffers,
> > because the hardware needs the full buffer list before starting a
> > decode job.
> > Am I free to change this queue parameter later, or is m2m_queue_init
> > the only place to do it ?
>
> It has to be set before the VIDIOC_STREAMON. After that you cannot
> change it anymore.
>
> But I don't think this is all that relevant, since this is something
> that the job_ready() callback should take care of. min_buffers_needed
> is really for hardware where the DMA engine cannot start unless that
> many buffers are queued. But in that case the DMA runs continuously
> capturing video, whereas here these are jobs and the DMA is only
> started when you can actually execute a job.

After doing some testing, overriding min_buffers_needed in queue_setup
is what works best for me.

When doing the initialization in start_streaming, the complete buffer
list needs to be configured in HW. The firmware can then choose any
buffer from this list during decoding later on.

I do this initialization by iterating with v4l2_m2m_for_each_dst_buf,
which requires all CAPTURE buffers to be queued in.

Cheers,
Maxime

> Regards,
>
>         Hans