[PATCH 2/4] spi: split up spi_new_device() to allow two stage registration.

[Grant Likely]

On Sat, May 24, 2008 at 12:43 AM, Grant Likely
<grant.likely at secretlab.ca> wrote:
> On Wed, May 21, 2008 at 6:17 PM, David Brownell <david-b at pacbell.net> wrote:
>> On Friday 16 May 2008, Grant Likely wrote:
>>>
>>> This patch splits the allocation and registration portions of code out
>>> of spi_new_device() and creates three new functions; spi_alloc_device(),
>>> spi_register_device(), and spi_device_release().
>>
>> I have no problem with the first two, but why the last?
>>
>> If the devices are always allocated by spi_alloc_device() as
>> they should be -- probably through an intermediary -- the
>> only public function necessary for that cleanup should be
>> the existing spi_dev_put().
>
> Ah, okay.  I'm still a bit fuzzy on the device model conventions.
> I'll remove that then.

I've dug into this some more.  spi_alloc_device only allocates the
memory.  It doesn't call device_initialize() to initialize the kref.
All of that behaviour is handled within device_register().  Therefore
if a driver uses spi_alloc_device() and then if a later part of the
initialization fails before spi_register_device() is called, then the
alloc'd memory needs to be freed, but spi_dev_put() won't work because
the kobj isn't set up so I need another function to handle freeing it
in on a failure path.

Should I switch things around to do device_initialize() in the alloc
function and call device_add() instead of device_register() in the
spi_register_device() function?  Is that sufficient to make
put_device() work?

Cheers,
g.

-- 
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.