On Wed, May 09, 2018 at 04:06:32PM -0700, Steve Longerbeam wrote:
>
>
> On 05/08/2018 03:12 AM, Sakari Ailus wrote:
> > On Fri, Apr 20, 2018 at 10:12:33AM -0700, Steve Longerbeam wrote:
> > > Hi Sakari,
> > >
> > >
> > > On 04/20/2018 05:24 AM, Sakari Ailus wrote:
> > > > Hi Steve,
> > > >
> > > > Thanks for the patchset.
> > > >
> > > > On Tue, Mar 20, 2018 at 05:37:19PM -0700, Steve Longerbeam wrote:
> > > > > v4l2_async_notifier_add_subdev() adds an asd to the notifier. It
> > > > > checks
> > > > > that the asd's match_type is valid and that no other equivalent asd's
> > > > > have already been added to this notifier's asd list, or to other
> > > > > registered notifier's waiting or done lists, and increments
> > > > > num_subdevs.
> > > > >
> > > > > v4l2_async_notifier_add_subdev() does not make use of the notifier
> > > > > subdevs
> > > > > array, otherwise it would have to re-allocate the array every time the
> > > > > function was called. In place of the subdevs array, the function adds
> > > > > the asd to a new master asd_list. The function will return error with
> > > > > a
> > > > > WARN() if it is ever called with the subdevs array allocated.
> > > > >
> > > > > In v4l2_async_notifier_has_async_subdev(),
> > > > > __v4l2_async_notifier_register(),
> > > > > and v4l2_async_notifier_cleanup(), alternatively operate on the
> > > > > subdevs
> > > > > array or a non-empty notifier->asd_list.
> > > > I do agree with the approach, but this patch leaves the remaining users
> > > > of
> > > > the subdevs array in the notifier intact. Could you rework them to use
> > > > the
> > > > v4l2_async_notifier_add_subdev() as well?
> > > >
> > > > There seem to be just a few of them --- exynos4-is and rcar_drif.
> > > I count more than a few :)
> > >
> > > % cd drivers/media && grep -l -r --include "*.[ch]"
> > > 'notifier[\.\-]>*subdevs[ ]*='
> > > v4l2-core/v4l2-async.c
> > > platform/pxa_camera.c
> > > platform/ti-vpe/cal.c
> > > platform/exynos4-is/media-dev.c
> > > platform/qcom/camss-8x16/camss.c
> > > platform/soc_camera/soc_camera.c
> > > platform/atmel/atmel-isi.c
> > > platform/atmel/atmel-isc.c
> > > platform/stm32/stm32-dcmi.c
> > > platform/davinci/vpif_capture.c
> > > platform/davinci/vpif_display.c
> > > platform/renesas-ceu.c
> > > platform/am437x/am437x-vpfe.c
> > > platform/xilinx/xilinx-vipp.c
> > > platform/rcar_drif.c
> > >
> > >
> > > So not including v4l2-core, the list is:
> > >
> > > pxa_camera
> > > ti-vpe
> > > exynos4-is
> > > qcom
> > > soc_camera
> > > atmel
> > > stm32
> > > davinci
> > > renesas-ceu
> > > am437x
> > > xilinx
> > > rcar_drif
> > >
> > >
> > > Given such a large list of the users of the notifier->subdevs array,
> > > I think this should be done is two steps: submit this patchset first,
> > > that keeps the backward compatibility, and then a subsequent patchset
> > > that converts all drivers to use v4l2_async_notifier_add_subdev(), at
> > > which point the subdevs array can be removed from struct
> > > v4l2_async_notifier.
> > >
> > > Or, do you still think this should be done all at once? I could add a
> > > large patch to this patchset that does the conversion and removes
> > > the subdevs array.
> > Sorry for the delay. I grepped for this, too, but I guess I ended up using
> > an expression that missed most of the users. :-(
> >
> > I think it'd be very good to perform that conversion --- the code in the
> > framework would be quite a bit cleaner and easier to maintain whereas the
> > per-driver conversions seem pretty simple, such as this on in
> > drivers/media/platform/atmel/atmel-isi.c :
> >
> > /* Register the subdevices notifier. */
> > subdevs = devm_kzalloc(isi->dev, sizeof(*subdevs), GFP_KERNEL);
> > if (!subdevs) {
> > of_node_put(isi->entity.node);
> > return -ENOMEM;
> > }
> >
> > subdevs[0] = &isi->entity.asd;
> >
> > isi->notifier.subdevs = subdevs;
> > isi->notifier.num_subdevs = 1;
> > isi->notifier.ops = &isi_graph_notify_ops;
>
>
> Yes, the conversions are fairly straightforward. I've completed that work,
> but it was a very manual conversion, every platform is different and needed
> careful study.
>
> Although I was careful about getting the error-out paths correct, there
> could
> be mistakes there, which would result in memory leaks. And obviously I can't
> re-test all these platforms. So this patch is very high-risk. More eyes are
> needed on it, ideally the maintainer(s) of each affected platform.
>
> I just submitted v4 of this series, but did not include this large un-tested
> patch in v4 for those reasons.
>
> Instead, this patch, and follow-up patches that strips support for subdevs
> array altogether from v4l2-async.c, and updates rst docs, are available at
> my
> media-tree mirror on github:
>
> g...@github.com:slongerbeam/mediatree.git
>
> in the branch 'remove-subdevs-array'. The branch is based off this series
> (branch 'imx-subdev-notifiers.6').
Would you be able to post these to the list? I'd really like this being
done as part of the related patchset, rather than leaving the mess in the
framework.
--
Sakari Ailus
sakari.ai...@linux.intel.com
