Re: [RFC PATCH 1/2] serial: 8250: Add a notifier chain for driver registration.

[David Daney <ddaney@caviumnetworks.com>]

On 03/17/2011 05:18 AM, Alan Cox wrote:
> On Wed, 16 Mar 2011 18:26:06 -0700
> David Daney<ddaney@caviumnetworks.com>  wrote:
>
>> The 8250 driver is a bit weird in that in addition to supporting
>> platform devices, extra devices can be added by calling
>> serial8250_register_port().
>>
>> The problem is that if we call serial8250_register_port() before the
>> driver is initialized Bad Things happen (we dereference NULL
>> pointers).
>>
>> There doesn't seem to be a general way to know if a driver has been
>> initialized
>
> I've had a bigger dig into this. I think the correct answer is probably
> "always go via platform devices or similar". That *is* the notifier in
> the kernel of today. serial8250_register_port ultimately should I think
> ultimatly become an internal helper.
>

That was kind of my thought too.  However we have all sorts of things 
calling serial8250_register_port(). Things like:

8250_pci.c
of_serial.c
8250_acorn.c
8250_gsc.c
.
.
.

The resulting view of the drivers in sysfs is that the little stub code 
that calls serial8250_register_port() is shown as the driver rather than 
serial8250.  But I suppose that is a matter of aesthetics more than 
function.

All those 'stub drivers' are relying on the ordering of module_init 
calls caused indirectly by the Makefile layout.  The path of least 
resistance is your suggestion that I use late_initcall() in my driver 
stub.  I actually tried that before hacking up this patch, but didn't 
like the idea of relying on *_initcall() ordering being necessary for 
correct initialization.

David Daney