Hi Hans,
On Friday, 8 December 2017 10:46:34 EET Hans Verkuil wrote:
> On 12/08/2017 08:54 AM, Laurent Pinchart wrote:
> > On Friday, 8 December 2017 03:08:17 EET Niklas Söderlund wrote:
> >> If the video device was registered by the complete() callback it should
> >> be unregistered when the driver is removed.
> >
> > The .remove() operation indicates device removal, not driver removal (or,
> > the be more precise, it indicates that the device is unbound from the
> > driver). I'd update the commit message accordingly.
> >
> >> Protect from printing an uninitialized video device node name by adding a
> >> check in rvin_v4l2_unregister() to identify that the video device is
> >> registered.
> >>
> >> Signed-off-by: Niklas Söderlund <niklas.soderlund+rene...@ragnatech.se>
> >> Reviewed-by: Kieran Bingham <kieran.bingham+rene...@ideasonboard.com>
> >> Reviewed-by: Hans Verkuil <hans.verk...@cisco.com>
> >> ---
> >>
> >> drivers/media/platform/rcar-vin/rcar-core.c | 2 ++
> >> drivers/media/platform/rcar-vin/rcar-v4l2.c | 3 +++
> >> 2 files changed, 5 insertions(+)
> >>
> >> diff --git a/drivers/media/platform/rcar-vin/rcar-core.c
> >> b/drivers/media/platform/rcar-vin/rcar-core.c index
> >> f7a4c21909da6923..6d99542ec74b49a7 100644
> >> --- a/drivers/media/platform/rcar-vin/rcar-core.c
> >> +++ b/drivers/media/platform/rcar-vin/rcar-core.c
> >> @@ -272,6 +272,8 @@ static int rcar_vin_remove(struct platform_device
> >> *pdev)>>
> >> pm_runtime_disable(&pdev->dev);
> >>
> >> + rvin_v4l2_unregister(vin);
> >
> > Unless I'm mistaken, you're unregistering the video device both here and
> > in the unbound() function. That's messy, but it's not really your fault,
> > the V4L2 core is very messy in the first place, and registering video
> > devices in the complete() handler is a bad idea. As that can't be fixed
> > for now,
> >
> > Acked-by: Laurent Pinchart <laurent.pinch...@ideasonboard.com>
> >
> > Hans, I still would like to hear your opinion on how this should be
> > solved. You've voiced a few weeks ago that register video devices at
> > probe() time isn't a good idea but you've never explained how we should
> > fix the problem. I still firmly believe that video devices should be
> > registered at probe time, and we need to reach an agreement on a technical
> > solution to this problem.
>
> I have tentatively planned to look into this next week. What will very
> likely have to happen is that we need to split off allocation from the
> registration, just as is done in most other subsystems. Allocation can be
> done at probe time, but the final registration step should likely be in the
> complete().
>
> To what extent that will resolve this specific issue I don't know. It will
> take me time to understand this in more detail.
I believe that splitting initialization from registration is a good idea, but
I still believe that video nodes should be registered at probe time
nonetheless. Let's discuss it again after next week when you'll have had time
to think about it.
> >> v4l2_async_notifier_unregister(&vin->notifier);
> >> v4l2_async_notifier_cleanup(&vin->notifier);
> >>
> >> diff --git a/drivers/media/platform/rcar-vin/rcar-v4l2.c
> >> b/drivers/media/platform/rcar-vin/rcar-v4l2.c index
> >> 178aecc94962abe2..32a658214f48fa49 100644
> >> --- a/drivers/media/platform/rcar-vin/rcar-v4l2.c
> >> +++ b/drivers/media/platform/rcar-vin/rcar-v4l2.c
> >> @@ -841,6 +841,9 @@ static const struct v4l2_file_operations rvin_fops =
> >> {
> >>
> >> void rvin_v4l2_unregister(struct rvin_dev *vin)
> >> {
> >> + if (!video_is_registered(&vin->vdev))
> >> + return;
> >> +
> >>
> >> v4l2_info(&vin->v4l2_dev, "Removing %s\n",
> >> video_device_node_name(&vin->vdev));
--
Regards,
Laurent Pinchart
