RE: [PATCH v2 1/2] Input: exc3000 - Simplify probe()

[Biju Das]

Hi Dmitry,

Thanks for the feedback.

> Subject: Re: [PATCH v2 1/2] Input: exc3000 - Simplify probe()
> 
> On Sat, Jul 22, 2023 at 05:51:17PM +0000, Biju Das wrote:
> > Hi Dmitry Torokhov,
> >
> > Thanks for the feedback.
> >
> > > Subject: Re: [PATCH v2 1/2] Input: exc3000 - Simplify probe()
> > >
> > > On Wed, Jul 19, 2023 at 06:43:47AM +0000, Biju Das wrote:
> > > > Hi Dmitry Torokhov,
> > > >
> > > > Thanks for the feedback.
> > > >
> > > > > Subject: Re: [PATCH v2 1/2] Input: exc3000 - Simplify probe()
> > > > >
> > > > > On Mon, Jul 17, 2023 at 06:45:27PM +0000, Biju Das wrote:
> > > > > > Hi Dmitry,
> > > > > >
> > > > > > > Subject: Re: [PATCH v2 1/2] Input: exc3000 - Simplify
> > > > > > > probe()
> > > > > > >
> > > > > > > On Mon, Jul 17, 2023 at 07:15:50PM +0100, Mark Brown wrote:
> > > > > > > > On Mon, Jul 17, 2023 at 04:35:02PM +0000, Biju Das wrote:
> > > > > > > >
> > > > > > > > > The .device_get_match_data callbacks are missing for I2C
> > > > > > > > > and SPI bus
> > > > > > > subsystems.
> > > > > > > > > Can you please throw some lights on this?
> > > > > > > >
> > > > > > > > It's the first time I've ever heard of that callback, I
> > > > > > > > don't know why whoever added it wouldn't have done those
> > > > > > > > buses in particular or if it just didn't happen.  Try
> > > > > > > > adding it and if it works send
> > > > > the patches?
> > > > > > >
> > > > > > > I think there is a disconnect. Right now
> > > > > > > device_get_match_data callbacks are part of
> > > > > > > fwnode_operations. I was proposing to add another optional
> > > > > > > device_get_match_data callback to 'struct
> > > bus_type'
> > > > > > > to allow individual buses control how match data is handled,
> > > > > > > before (or after) jumping into the fwnode-backed
> > > > > > > device_get_match_data
> > > > > callbacks.
> > > > > >
> > > > > > That is what implemented here [1] and [2] right?
> > > > > >
> > > > > >
> > > > > > First it check for fwnode-backed device_get_match_data
> > > > > > callbacks and Fallback is bus-type based match.
> > > > > >
> > > > > > Looks like you are proposing to unify [1] and [2] and you want
> > > > > > the logic to be other way around. ie, first bus-type match,
> > > > > > then fwnode-backed callbacks?
> > > > > >
> > > > >
> > > > > I do not have a strong preference for the ordering, i.e. I think
> > > > > it is perfectly fine to do the generic fwnode-based lookup and
> > > > > if there is no match have bus method called as a fallback,
> > > >
> > > > That involves a bit of work.
> > > >
> > > > const void *device_get_match_data(const struct device *dev);
> > > >
> > > > const struct i2c_device_id *i2c_match_id(const struct
> > > > i2c_device_id
> > > *id,
> > > >                                    const struct i2c_client
> > > > *client);
> > > >
> > > > const struct spi_device_id *spi_get_device_id(const struct
> > > > spi_device *sdev);
> > > >
> > > > Basically, the bus-client driver(such as exc3000) needs to pass
> > > > struct device and device_get_match_data after generic fwnode-based
> > > > lookup, needs to find the bus type based on struct device and call
> > > > a new generic
> > > > void* bus_get_match_data(void*) callback, so that each bus
> > > > interface can do a match.
> > >
> > > Yes, something like this (which does not seem that involved to
> me...):
> >
> > Looks it will work.
> >
> > But there is some 2 additional checks in core code, every driver which
> is not bus type need to go through this checks.
> >
> > Also in Bus specific callback, there are 2 additional checks.
> >
> > So, performance wise [1] is better.
> 
> I do not believe this is a concern whatsoever: majority of
> architectures/boards have been converted to ACPI/DT, which are being
> matched first as they are now, so the fallback to bus-specific matching
> against bus-specific device ID tables will be very infrequent.
> Additionally, device_get_match_data() is predominantly called from
> driver probe paths, so we need not be concerned with it being used with
> class devices or other kinds of devices not associated with a bus.

Looks like most of the i2c client driver uses similar handling for 
ACPI/DT and ID tables. If that is the case, it is good to have this
proposed change which will simplify most of the drivers listed in [1]

[1] https://elixir.bootlin.com/linux/latest/A/ident/i2c_match_id

Eg: drivers/hwmon/pmbus/ibm-cffps.c

	enum versions vs = cffps_unknown;
	const void *md = of_device_get_match_data(&client->dev);
	const struct i2c_device_id *id;

	if (md) {
		vs = (enum versions)md;
	} else {
		id = i2c_match_id(ibm_cffps_id, client);
		if (id)
			vs = (enum versions)id->driver_data;
	}

The above code can be converted to 
     vs = (enum versions)device_get_match_data(&client->dev);

> 
> >
> > Moreover, we need to avoid code duplication with [1]
> >
> > [1]
> 
> If and when my proposed solution gets into the kernel we can drop
> i2c_get_match_data() altogether.

Agreed. Will wait for other people's view on this topic.

Cheers,
Biju