Re: [PATCH 1/7] driver: introduce less error-prone dev_get_drvdata alternative

[Ahmad Fatoum <a.fatoum@pengutronix.de>]

Hello Sascha,

On 9/29/20 9:32 AM, Sascha Hauer wrote:
> On Mon, Sep 28, 2020 at 05:42:41PM +0200, Ahmad Fatoum wrote:
>> We use dev_get_drvdata to get the driver match data associated with a
>> device. This has two shortcomings:
>>
>>  - Linux has dev_get_drvdata too, which returns a private pointer for
>>    driver specific info to associate with a device. We use dev->priv
>>    (or more often container_of) for that in barebox instead
>>
>>  - It nearly always involves a cast to a double pointer, which is
>>    error-prone as size and alignment match need to be ensured
>>    on the programmer's part and can easily be gotten wrong:
>> 	enum dev_type type
>> 	dev_get_drvdata(dev, (const void **)&type);
>>
>> Add a new function that instead of using a double pointer argument,
>> returns the pointer directly:
>>
>>  - For normal pointer driver data, no cast is necessary
>>  - For integer driver data casted to a pointer for storage,
>>    the cast is still necessary, but
>>
>> Signed-off-by: Ahmad Fatoum <a.fatoum@pengutronix.de>
>> ---
>>  drivers/base/driver.c | 11 +++++++++++
>>  include/driver.h      | 18 ++++++++++++++++++
>>  2 files changed, 29 insertions(+)
>>
>> diff --git a/drivers/base/driver.c b/drivers/base/driver.c
>> index 412db6c40699..3205bbc3c33b 100644
>> --- a/drivers/base/driver.c
>> +++ b/drivers/base/driver.c
>> @@ -500,3 +500,14 @@ int dev_get_drvdata(struct device_d *dev, const void **data)
>>  
>>  	return -ENODEV;
>>  }
>> +
>> +const void *device_get_match_data(struct device_d *dev)
>> +{
>> +	if (dev->of_id_entry)
>> +		return dev->of_id_entry->data;
>> +
>> +	if (dev->id_entry)
>> +		return (void *)dev->id_entry->driver_data;
>> +
>> +	return NULL;
>> +}
> 
> Can we please give the caller a possibility to distinguish between
> errors and a valid NULL pointer?

The only error possible is that we matched the device by name
and not of_compatible or platform id entry.

If the match data is a valid pointer:
 -> It doesn't matter, why we have no match data either way.

If the match data is a casted integer (e.g. enum):
 The driver author should either:
   -> place the default enum value as first one,
      so no match data => default
   -> should add an initial DEVICE_TYPE_UNKNOWN = 0 in the enum
      and handle it appropriately

I like the function signature like that, I don't really see
a need to adjust it.

> As you realize in your series some drivers cast the returned value into
> an integer type and use 0 as a valid possibility. These need an extra
> explanation why we can accept that case. I can think of different
> possibilies to get that straight:
> 
> - Put a real pointer into matchdata. This is really preferred as it
>   motivates people to put flags into a (struct type) matchdata which
>   doesn't lead to excessive if (type == foo || type == bar || type ==
>   baz) we sometimes see in drivers.

We have a real pointer there already. The problem is migrating the
existing drivers.

> - Return an ERR_PTR from device_get_match_data(). this is less likely
>   interpreted as a valid int value

Doesn't cover all cases. Also for the normal use, it means
you need to have to check with IS_ERR_OR_NULL everywhere to
be sure you don't dereference a NULL pointer.

> - add a int * ret argument to device_get_match_data() which returns
>   the error code.

Would work, but my preference is leaving it to the drivers.
The drivers need to handle errors anyway.

> 
> Sascha
> 

-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |

_______________________________________________
barebox mailing list
barebox@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/barebox