Re: [PATCH v8 4/8] media: platform: Add Cedrus VPU decoder driver

[Tomasz Figa]

On Thu, Sep 6, 2018 at 4:39 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>
> On 09/06/2018 09:25 AM, Tomasz Figa wrote:
> > On Thu, Sep 6, 2018 at 4:01 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
> >>
> >> On 09/05/2018 06:29 PM, Paul Kocialkowski wrote:
> >>> Hi and thanks for the review!
> >>>
> >>> Le lundi 03 septembre 2018 Ã 11:11 +0200, Hans Verkuil a Ãcrit :
> >>>> On 08/28/2018 09:34 AM, Paul Kocialkowski wrote:
> >>>>> +static int cedrus_request_validate(struct media_request *req)
> >>>>> +{
> >>>>> +   struct media_request_object *obj, *obj_safe;
> >>>>> +   struct v4l2_ctrl_handler *parent_hdl, *hdl;
> >>>>> +   struct cedrus_ctx *ctx = NULL;
> >>>>> +   struct v4l2_ctrl *ctrl_test;
> >>>>> +   unsigned int i;
> >>>>> +
> >>>>> +   list_for_each_entry_safe(obj, obj_safe, &req->objects, list) {
> >>>>
> >>>> You don't need to use the _safe variant during validation.
> >>>
> >>> Okay, I'll use the regular one then.
> >>>
> >>>>> +           struct vb2_buffer *vb;
> >>>>> +
> >>>>> +           if (vb2_request_object_is_buffer(obj)) {
> >>>>> +                   vb = container_of(obj, struct vb2_buffer, req_obj);
> >>>>> +                   ctx = vb2_get_drv_priv(vb->vb2_queue);
> >>>>> +
> >>>>> +                   break;
> >>>>> +           }
> >>>>> +   }
> >>>>
> >>>> Interesting question: what happens if more than one buffer is queued in the
> >>>> request? This is allowed by the request API and in that case the associated
> >>>> controls in the request apply to all queued buffers.
> >>>>
> >>>> Would this make sense at all for this driver? If not, then you need to
> >>>> check here if there is more than one buffer in the request and document in
> >>>> the spec that this is not allowed.
> >>>
> >>> Well, our driver was written with the (unformal) assumption that we
> >>> only deal with a pair of one output and one capture buffer. So I will
> >>> add a check for this at request validation time and document it in the
> >>> spec. Should that be part of the MPEG-2 PIXFMT documentation (and
> >>> duplicated for future formats we add support for)?
> >>
> >> Can you make a patch for vb2_request_has_buffers() in videobuf2-core.c
> >> renaming it to vb2_request_buffer_cnt() and returning the number of buffers
> >> in the request?
> >>
> >> Then you can call it here to check that you have only one buffer.
> >>
> >> And this has to be documented with the PIXFMT.
> >>
> >> Multiple buffers are certainly possible in non-codec scenarios (vim2m and
> >> vivid happily accept that), so this is an exception that should be
> >> documented and checked in the codec driver.
> >
> > Hmm, isn't it still 1 buffer per 1 queue and just multiple queues
> > included in the request?
>
> No. The request API allows multiple buffers for the same vb2_queue to be
> queued for the same request (obviously when the request is committed, the
> buffers are queued to the driver in the same order).
>
> >
> > If we indeed allow multiple buffers for the same queue in a request,
> > we shouldn't restrict this on a per-driver basis. It's definitely not
> > a hardware limitation, since the driver could just do the same as if 2
> > requests with the same controls were given.
>
> That's how it operates: for all buffers in the request the same controls
> apply. But does this make sense for codecs? If the control(s) with the
> codec metadata always change for every buffer, then having more than one
> buffer in the request is senseless and the driver should check for this
> in the validation step.
>
> If it *does* make sense in some circumstances to have the same metadata
> for multiple buffers, then it should be checked if the cedrus driver
> handles this correctly.

Just FYI, we may want to move this discussion to Alex's RFC with
documentation of stateless interface:
https://patchwork.kernel.org/patch/10583233/

Best regards,
Tomasz