[PATCH 2/3 v8] mailbox: Enable BCM2835 mailbox support

[swarren@wwwdotorg.org (Stephen Warren)]

On 05/08/2015 01:19 PM, Eric Anholt wrote:
> Alexander Stein <alexanders83@web.de> writes:
> 
>> On Thursday 07 May 2015, 12:54:20 wrote Eric Anholt:
>>> Noralf Tr?nnes <noralf@tronnes.org> writes:
>>> 
>>>> Den 05.05.2015 22:27, skrev Eric Anholt:
>>>>> From: Lubomir Rintel <lkundrak@v3.sk>
>>>>> 
>>>>> This mailbox driver provides a single mailbox channel to
>>>>> write 32-bit values to the VPU and get a 32-bit response.
>>>>> The Raspberry Pi firmware uses this mailbox channel to
>>>>> implement firmware calls, while Roku 2 (despite being
>>>>> derived from the same firmware tree) doesn't.
>>>>> 
>>>>> The driver was originally submitted by Lubomir, based on
>>>>> the out-of-tree 2708 mailbox driver.  Eric Anholt fixed it
>>>>> up for upstreaming, with the major functional change being
>>>>> that it now has no notion of multiple channels (since that
>>>>> is a firmware-dependent concept) and instead the
>>>>> raspberrypi-firmware driver will do that bit-twiddling in
>>>>> its own messages.
>>>> ...
>>>>> +static struct platform_driver bcm2835_mbox_driver = { +
>>>>> .driver = { +		.name = "bcm2835-mbox", +		.owner =
>>>>> THIS_MODULE, +		.of_match_table = bcm2835_mbox_of_match, +
>>>>> }, +	.probe		= bcm2835_mbox_probe, +	.remove		=
>>>>> bcm2835_mbox_remove, +}; 
>>>>> +module_platform_driver(bcm2835_mbox_driver);
>>>> 
>>>> I have tested this driver and the firmware driver booting
>>>> directly from the VideoCore bootloader (no uboot). The
>>>> mailbox driver loads too late to turn on USB power:
>>> 
>>> Yeah, I have a patch on my branches that returns -EPROBE_DEFER
>>> when trying to get a power domain and not finding the provider.
>>> It was rejected by the maintainers in favor of a proposed
>>> solution whose description I didn't quite follow.
>> 
>> Do you have a link for this thread?
> 
> https://lkml.org/lkml/2015/3/11/483

That's really odd; -EPROBE_DEFER was clearly invented exactly to
handle dependencies just like this. Playing with initcall levels
simply isn't scalable, and in the main people are actively working not
to use them for dependencies like this; they're far too implicit.
While the timeout mentioned earlier in the thread might work (I didn't
really look at the details), again it's far too indirect/accidental to
be a good solution.