Re: [PATCH v5 15/16] leds: Add common LED binding parsing support to LED class/core

From: Vaittinen, Matti
Date: Wed Nov 20 2019 - 02:31:21 EST



On Tue, 2019-11-19 at 20:30 +0100, Jacek Anaszewski wrote:
> Hi Matti,
>
> On 11/19/19 8:21 AM, Vaittinen, Matti wrote:
> > Hello Jacek,
> >
> > On Mon, 2019-11-18 at 22:55 +0100, Jacek Anaszewski wrote:
> > > Hi Matti,
> > >
> > > Thank you for the patch. If your driver does not depend
> > > on it then please send is separately.
> >
> > The BD71828 depends on device-tree node look-up. It does not
> > utilize
> > the common property parsing. I could of course do the child dt-node
> > walking in BD71828 driver - but it kind of breaks my motivation to
> > do
> > the LED core improvement if I anyways need to do the parsing in
> > BD71828
> > driver ;)
>
> If you do not plan on spending too much time on contributing this
> set then I propose adhering to the currently used parsing schema :-)

I have no objections on doing few iterations of the patches. And I tend
to take care of problems my changes cause. So I am prepared to spend
the required time fixing the node look-up and common property parsing
for drivers I do break. What I am not prepared is to change and test
all of the existing drivers - so it's better to not promise such :)

> And you have to know that from this development cycle I handed
> over LED tree maintenance to Pavel Machek, so you will require
> to have his acceptance in the first place.

Well, then I for sure wait for Pavel's take on this. In general I have
had some positive feedback about doing the DT node look-up and common
property parsing in a centralized manner in LED core. So maybe also
Pavel sees the value of adding this now for new drivers - instead of
adding one more driver with copy-paste node look-up code. I want to
thank you for all the comments though, it's nice that you have been
active on this topic!

> > > Besides, we would require
> > > to convert many of current LED drivers to verify how the
> > > proposed parsing mechanism will work with them.
> >
> > I see the risk you are pointing out. And I actually think we could
> > default to old mechanism if of_match or match_property is not given
> > (for now). I could then see the existing drivers who use init_data
> > -
> > and ensure those are initializing the new match_property and
> > of_match
> > in init_data with 0. That would be quite trivial task.
> >
> > That would allow us to convert and test existing drivers one-by-one
> > while allowing new drivers to offload the LED node look-up and
> > common
> > property parsing to LED core. No risk, but less drivers to convert
> > in
> > the future - and simpler drivers to maintain when all of them do
> > not
> > need to duplicate node look-up or basic property parsing ;)
>
> I personally would prefer to do the massive driver update to using
> the new mechanism. I know that this is time consuming but we are not
> in a hurry.

I understand the preference of massive update - but I also know that if
we wait for someone to do a massive update and neglect improvements
done in small steps, then there is a risk that there won't be any
updates at all...

> > To make this more concrete:
> >
> > We can only do the new DT node look-up if either
> > if (init_data->match_property.name && init_data-
> > >match_property.size)
> > or
> > if (init_data->of_match)
> > That would keep the node-lookup same for all existing drivers.
> >
> > Eg,
> > led_find_fwnode could for now just do:
> >
> > struct fwnode_handle *led_find_fwnode(struct device *parent,
> > struct led_init_data *init_data)
> > {
> > /*
> > * This should never be called W/O init data.
> > */
> > if (!init_data)
> > return NULL;
> >
> > /*
> > * For old drivers which do not populate new match information
> > * just directly use the given init_data->fwnode no matter if
> > * it is NULL or not. - as old functionality did.
> > */
> > if ( (!init_data->match_property.name ||
> > !init_data->match_property.size) && !init_data->of_match)
> > return init_data->fwnode;
> >
> > /* match information was given - do node look-up */
> > ...
> > }
> >
> > Furthermore, the common property parsing could also be (for now)
> > done
> > only if match data is given:
> >
> > /*
> > * For now only allow core to parse DT properties if
> > * parsing is explicitly requested by driver or if core has
> > * found new match data from init_data and then set the flag
> > */
> > if (INVENT_A_COOL_NEW_FLAG_NAME_HERE)
> > led_add_props(led_cdev, &props);
> >
> > or just simply:
> > if ((init_data->match_property.name &&
> > init_data->match_property.size) || init_data->of_match)
> > led_add_props(led_cdev, &props);
> >
> > (but this won't allow driver to ask for common parsing even if it
> > was
> > verified for this drv to work - hence I like the flag better)
> >
> > And if you don't feel confident I can even drop the "common
> > property
> > parsing" from the series and leave only the "node look-up if match-
> > data
> > was given" to it.
> >
> > Anyways, I would like to introduce this support while I am working
> > with
> > the BD71828 driver which really has the LEDs - but I can modify the
> > patch series so that it only impacts to drivers which implement the
> > new
> > match data in init_data and leave old drivers to be converted one-
> > by-
> > one when they can be tested.
> >
> > > I've been testing
> > > my LED name composition series using QEMU and stubbing things in
> > > drivers where necessary and I propose to use the same approach
> > > in this case.
> >
> > I don't plan to do any mass-conversion as it is somewhat risky. I
> > can
>
> You do not need hardware to test DT parsing as I mentioned before,
> so I don't see too much risk involved.

Yes - if you have the time to test all the drivers at once - and
assuming you don't do some silly mistake there. Final verification
should always be done in HW. But as I said, I want to ensure all the
drivers I convert to new mechanism will work (the best I can) - and as
I can't test all the drivers I won't do mass-conversion. I have offered
a initial solution in the patch (and suggested reduced version to
mitigate the risk of breaking anything in this email) - and I see this
beneficial and as a good starting point enabling the rest of the
improvements. But as you said, we need also Pavel's take on this.

> > do conversion to some of the drivers (simple ones which I can
> > understand without too much of pain) - and ask if anyone having
> > access
> > to actual HW with LEDs could be kind enough to test the patch for
> > the
> > device. Tested drivers can then be taken in-tree as examples. And
> > who
> > knows, maybe there is some developers looking for a hobby project
> > with
> > access to LED controller to help with the rest ;) I don't have the
> > ambition to change all of the LED drivers but I think I can give my
> > 10
> > cents by contributing the mechanism and doing few examples :)
>
> If you want to introduce good, robust mechanism, then it should be
> tested against widest possible spectrum of use cases.

Yes. OTOH, if the mechanism is sub-optimal, then the beauty of open
source is that it can be improved. Preparing in advance for something
that never happens is also a waste. But I guess we don't need to
discuss this philosphy here :)

> > Anyways, please let me know if you think you could accept patch
> > which
> > won't change existing driver functionality - but allows new drivers
> > to
> > not duplicate the code. Else I'll just duplicate the lookup code in
> > one
> > more driver and hope I don't have another PMIC with LED controller
> > on
> > my table too soon...
> >
> > (I am having "some" pressure to do few other tasks I recently got.
> > So I
> > am afraid I won't have too much time to invest on LEDs this year :(
> > Thus setting up the qemu and starting with stubbing is really not
> > an
> > option for me at this phase).
>
> As mentioned before - I no longer apply patches so you will need to
> consult Pavel, but I bet he will have similar opinion.

Who knows, maybe he can see this differently :) Thanks anyways!

Br,
Matti Vaittinen