[PATCH RFC] usb gadget: introduce usb_gadget_probe_driver

[Uwe Kleine-König]

On Fri, Jul 30, 2010 at 06:08:23PM +0200, Michał Nazarewicz wrote:
> On Fri, 30 Jul 2010 17:26:02 +0200, Uwe Kleine-König <u.kleine-koenig at pengutronix.de> wrote:
>> On Fri, Jul 30, 2010 at 05:16:46PM +0200, Michał Nazarewicz wrote:
>>> On Fri, 30 Jul 2010 16:49:14 +0200, Uwe Kleine-König <u.kleine-koenig at pengutronix.de> wrote:
>>>
>>>> This is like usb_gadget_register_driver with the only difference that it
>>>> gets the bind function as parameter instead of using driver->bind.  This
>>>> allows fixing section mismatches like
>>>>
>>>> 	WARNING: drivers/usb/gadget/g_printer.o(.data+0xc): Section mismatch in
>>>> 	reference from the variable printer_driver to the function
>>>> 	.init.text:printer_bind()
>>>> 	The variable printer_driver references
>>>> 	the function __init printer_bind()
>>>>
>>>> by using usb_gadget_probe_driver with driver->bind = NULL.  When all
>>>> drivers are fixed to use the new function the bind member of struct
>>>> usb_gadget_driver can go away.
>>>>
>>>> Signed-off-by: Uwe Kleine-König <u.kleine-koenig at pengutronix.de>
>>>> Cc: Michal Nazarewicz <m.nazarewicz at samsung.com>
>>>> Cc: Greg Kroah-Hartman <gregkh at suse.de>
>
> BTW. Dunno if there is a reason to put me on Cc and Greg will be in
> Signed-off-by anyway.
>
>>>> ---
>>>> Hello,
>>>>
>>>> there is an alternative patch in Greg's tree [1], but IMHO mine is
>>>> better as it doesn't need __ref.
>>>>
>>>> Thoughts?
>>>
>>> Personally I don't see advantage of this over changing the __init to __ref.
>>> Or am I missing something?  I see your patch as an unnecessary API change.
>>> The way I understand it, __ref was introduced exactly for cases like this
>>> one where function is referenced from "normal" data but used only during
>>> init.  Could you try to clarify for me why you think your solution is
>>> better then mine?
>
>> - Using __ref instead of __init moves all bind functions from .init.text
>>   to .text and so the code isn't freed after booting or module loading.
>>   (OK, you could fix this by marking the driver structs as __ref and
>>   keep __init for the bind functions.)
>
> I believe this to be untrue.  __ref puts code in .ref.text which AFAIK is
> freed.
I'm pretty sure it's not freed.  .ref.text is of course correct.

>> - Using __ref might hide section mismatches.  (Your patch hasn't this
>>   problem as the bind functions used to live in .init.text, so any
>>   reference to .init should be OK assuming that it was OK to live in
>>   .init.text in the first place.  But when marking the driver structs
>>   this is an issue.)  That's why I don't like __ref and said my patch
>>   were better as it doesn't make use of it.
>>
>> - The bind functions are only called at init time, so there is no need
>>   to save a pointer to it.
>
> OK, I see some merit in this approach.  However, the same issue is with
> usb_configuration and ?usb_composite_driver -- those should be changed
> in the same way or it would defeat the purpose of the patch.
I didn't skip these on purpose.  I just stumbled over the gadget drivers
first.

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | http://www.pengutronix.de/  |