The of_device_id structure uses a void pointer to associate additional
driver-private data with device id, most commonly used to refer to a
structure containing additional characteristics of a particular chip.
However i2c_device_id uses an unsigned long. While it can easily be
converted to a pointer, several drivers use it to store a scalar (and it
is possible to use a pointer in of_device_id to store a scalar value as
well). The worst case comes when OF part of a driver uses pointers,
while legacy part is using scalars, causing constructs like:

	data = device_get_match_data(...);
	if (!data)
		data = &some_array[i2c_match_id(...)->data];
	...

To avoid this introduce a const void "data" pointer to i2c_device_id as
well, so that we can convert drivers one by one, and drop driver_data
member in the end.

The end goal is to clean up all helpers and make device_get_match_data()
work universally for all ACPI, DT, and legacy variants.

Signed-off-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
---
 drivers/i2c/i2c-core-base.c     | 2 +-
 include/linux/mod_devicetable.h | 3 ++-
 2 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/i2c/i2c-core-base.c b/drivers/i2c/i2c-core-base.c
index 60746652fd52..75d12d79d8f1 100644
--- a/drivers/i2c/i2c-core-base.c
+++ b/drivers/i2c/i2c-core-base.c
@@ -126,7 +126,7 @@ const void *i2c_get_match_data(const struct i2c_client *client)
 		if (!match)
 			return NULL;
 
-		data = (const void *)match->driver_data;
+		data = match->data ?: (const void *)match->driver_data;
 	}
 
 	return data;
diff --git a/include/linux/mod_devicetable.h b/include/linux/mod_devicetable.h
index b0678b093cb2..21ac2a96e828 100644
--- a/include/linux/mod_devicetable.h
+++ b/include/linux/mod_devicetable.h
@@ -477,7 +477,8 @@ struct rpmsg_device_id {
 
 struct i2c_device_id {
 	char name[I2C_NAME_SIZE];
-	kernel_ulong_t driver_data;	/* Data private to the driver */
+	kernel_ulong_t driver_data;	/* This field is obsolete */
+	const void *data;		/* Data private to the driver */
 };
 
 /* pci_epf */

-- 
2.41.0.694.ge786442a9b-goog

On Mon, Aug 14, 2023 at 02:52:49PM -0700, Dmitry Torokhov wrote:
> The of_device_id structure uses a void pointer to associate additional
> driver-private data with device id, most commonly used to refer to a
> structure containing additional characteristics of a particular chip.
> However i2c_device_id uses an unsigned long. While it can easily be
> converted to a pointer, several drivers use it to store a scalar (and it
> is possible to use a pointer in of_device_id to store a scalar value as
> well). The worst case comes when OF part of a driver uses pointers,
> while legacy part is using scalars, causing constructs like:
> 
> 	data = device_get_match_data(...);
> 	if (!data)
> 		data = &some_array[i2c_match_id(...)->data];
> 	...
> 
> To avoid this introduce a const void "data" pointer to i2c_device_id as
> well, so that we can convert drivers one by one, and drop driver_data
> member in the end.
> 
> The end goal is to clean up all helpers and make device_get_match_data()
> work universally for all ACPI, DT, and legacy variants.

So, we have in the parallel the activity to make buses to have a callback,
why do we need this one? Looks to me as yet another 1000+ churn for not
much value. What the good outcome of this is constification, but maybe
we can find the way on how to prove const to stay over the kernel_ulong_t
transformations for all device ID tables?

-- 
With Best Regards,
Andy Shevchenko