Re: [RFCv4,19/21] media: vim2m: add request support

[Paul Kocialkowski]

Hi,

On Tue, 2018-03-13 at 19:24 +0900, Alexandre Courbot wrote:
> On Fri, Mar 9, 2018 at 11:35 PM, Paul Kocialkowski
> <paul.kocialkowski@xxxxxxxxxxx> wrote:
> > Hi,
> > 
> > On Thu, 2018-03-08 at 22:48 +0900, Alexandre Courbot wrote:
> > > Hi Paul!
> > > 
> > > Thanks a lot for taking the time to try this! I am also working on
> > > getting it to work with an actual driver, but you apparently found
> > > rough edges that I missed.
> > > 
> > > On Thu, Mar 8, 2018 at 1:37 AM, Paul Kocialkowski
> > > <paul.kocialkowski@xxxxxxxxxxx> wrote:
> > > > On Tue, 2018-02-20 at 13:44 +0900, Alexandre Courbot wrote:

[...]

> > > > > +static int vim2m_request_submit(struct media_request *req,
> > > > > +                             struct media_request_entity_data
> > > > > *_data)
> > > > > +{
> > > > > +     struct v4l2_request_entity_data *data;
> > > > > +
> > > > > +     data = to_v4l2_entity_data(_data);
> > > > 
> > > > We need to call v4l2_m2m_try_schedule here so that m2m
> > > > scheduling
> > > > can
> > > > happen when only 2 buffers were queued and no other action was
> > > > taken
> > > > from usespace. In that scenario, m2m scheduling currently
> > > > doesn't
> > > > happen.
> > > 
> > > I don't think I understand the sequence of events that results in
> > > v4l2_m2m_try_schedule() not being called. Do you mean something
> > > like:
> > > 
> > > *
> > > * QBUF on output queue with request set
> > > * QBUF on capture queue
> > > * SUBMIT_REQUEST
> > > 
> > > ?
> > > 
> > > The call to vb2_request_submit() right after should trigger
> > > v4l2_m2m_try_schedule(), since the buffers associated to the
> > > request
> > > will enter the vb2 queue and be passed to the m2m framework, which
> > > will then call v4l2_m2m_try_schedule(). Or maybe you are thinking
> > > about a different sequence of events?
> > 
> > This is indeed the sequence of events that I'm seeing, but the
> > scheduling call simply did not happen on vb2_request_submit. I
> > suppose I will need to investigate some more to find out exactly
> > why.
> > 
> > IIRC, the m2m qbuf function is called (and fails to schedule) when
> > the
> > ioctl happens, not when the task is submitted.
> > 
> > This issue is seen with vim2m as well as the rencently-submitted
> > sunxi-
> > cedrus driver (with the in-driver calls to v4l2_m2m_try_schedule
> > removed, obviously). If needs be, I could provide a standalone test
> > program to reproduce it.
> 
> If you have a standalone program that can reproduce this on vim2m,
> then I would like to see it indeed, if only to understand what I have
> missed.

You can find the test file for this use case at:
https://gist.github.com/paulkocialkowski/4cfa350e1bbe8e3bf714480bba83ea72

Cheers!

-- 
Paul Kocialkowski, Bootlin (formerly Free Electrons)
Embedded Linux and kernel engineering
https://bootlin.com

Attachment: signature.asc
Description: This is a digitally signed message part