Re: [PATCH 3/3] omap: add hwspinlock device

[Kevin Hilman <khilman@deeprootsystems.com>]

Ohad Ben-Cohen <ohad@wizery.com> writes:

> On Wed, Oct 20, 2010 at 1:53 AM, Kevin Hilman
> <khilman@deeprootsystems.com> wrote:
>>> And to allow early board code to reserve specific hwspinlock numbers
>>> for predefined use-cases, we probably want to be before arch_initcall.
>>
>> There's no reason for board code to have to do this at initcall time.
>
> If we want to have allow both allocations of predefined hwspinlocks
> with omap_hwspinlock_request_specific(int), and dynamic allocations
> (where we don't care about the specific instance of the hwspinlock we
> will get) with omap_hwspinlock_request(), we must ensure that the
> former _specific() API will never be called after the latter.
>
> If we will allow drivers to call omap_hwspinlock_request() before all
> callers of omap_hwspinlock_request_specific() completed, then things
> will break (because drivers might start getting hwspinlocks that are
> predefined for dedicated use cases on the system).
>
> So if we want the _specific API to work, we can only allow early board
> code to use it in order to reserve those predefined hwspinlocks before
> drivers get the chance to call omap_hwspinlock_request().
>
> The tempting alternative is not to provide the
> omap_hwspinlock_request_specific() API at all (which is something we
> discussed internally).
>
> Let's take the i2c-omap for example.
>
> It sounds like it must have a predefined hwspinlock, but what if:
>
> 1. It will use omap_hwspinlock_request() to dynamically allocate a hwspinlock
> 2. Obviously, the hwspinlock id number must be communicated to the M3
> BIOS, so the i2c-omap will publish that id using a small shared memory
> entry that will be allocated for this sole purpose
> 3. we will make sure that 1+2 completes before the remote processor is
> taken out of reset
>
> This does not require any smart IPC and it will allow us to get rid of
> the omap_hwspinlock_request_specific() API and its early-callers
> requirement.

Yes, that would indeed simplify things. 

> All we will be left to take care of is the order of the ->probe()
> execution (assuming we want both the i2c and the hwspinlock drivers to
> be device_initcall)

I understand the dependency between i2c and hwspinlock for some
platforms with a shared i2c bus.  Besides that being a broken hardware
design, I can see the need for the i2c driver to take a hwspinlock for
i2c xfers, but why does the i2c driver need to take the hwspinlock at
probe time?  Presumably, this is before the remote cores are executing
code.

>>
>> This kind of thing needs to be done by platform_data function pointers,
>> as is done for every other driver that needs platform-specific driver
>> customization.
>
> Why would we need platform-specific function pointers here ? I'm not
> sure I'm following this one.

So that board code (built-in) does not call the hwspinlock driver
(potentially a module.)

The way to solve this is to have platform_data with function pointers,
so that when the driver's ->probe() is done, it can call cany custom
hooks registered by the board code.

Kevin