[RFC 17/17] clk: zynq: remove call to of_clk_init

[Sebastian Hesselbarth]

On 08/23/13 19:19, Sören Brinkmann wrote:
> On Fri, Aug 23, 2013 at 11:30:18AM +0200, Sebastian Hesselbarth wrote:
>> On 08/23/13 02:59, Sören Brinkmann wrote:
>>> On Thu, Aug 22, 2013 at 05:26:47PM -0700, Sören Brinkmann wrote:
>>>> On Tue, Aug 20, 2013 at 04:04:31AM +0200, Sebastian Hesselbarth wrote:
>>>>> With arch/arm calling of_clk_init(NULL) from time_init(), we can now
>>>>> remove it from corresponding drivers/clk code.
>>>>
>>>> I think that would break Zynq.
>>>> If I see this correctly you call of_clk_init() from common code,
>>>> _before_ the SOC specific time init function is called.
>>>> The problem is, that we have code setting up a global pointer which is
>>>> required by zynq_clk_setup() which is triggered when of_clk_init() is
>>>> called.
> [ ... ]
>> thanks for looking into this. I also had a look at the files in
>> question. Based on Steffen's proposal, I prepared a diff that should do
>> the trick. It moves zynq_slcr_init() to early_init, instead of reusing
>> another hook that has magic cow powers (it calls irqchip_init that zynq
>> also wants sooner or later).
>>
>> Also, it removes zynq_clock_init() and let zynq_clk_setup() map the
>> register itself by finding the node and use of_iomap(). I realized that
>> clock registers are quite separated within slcr, so you can consider
>> to have your own node for the clk-provider. As Steffen is proposing
>> this but mentioned incompatible DT changes, I chose that intermediate
>> step above.
>>
>> It would be great, if you test the diff and prepare a patch out of
>> it, that I pick-up in the patch set. That way, we also have your
>> Signed-off on it.
>
> I looked into this. Looks like init_early() happens to early. I suspect
> slab is missing to make zynq_slcr_init() work. So, I moved it into
> init_irq(). Is there any init_call() type which is called at the correct
> time?

Sören,

I mistakenly assumed init_early is after mm, so of course my proposal
does not work as it should. I am fine with moving it to init_irq() until
you find the best solution (or until we have the same "mess" with
default init_irq hook).

> I looked briefly into syscon and regmap, and that does actually look
> promising and to really fix this mess, I guess we have to wait a little
> until Steffen finishes his work on it.

IIRC, both syscon and regmap will require you to have devices ready.
I haven't followed all recent discussions about early device
registration. Anyway, it will not help much in the current approach
to get rid of custom .init_timer and maybe .init_irq later.

> To facilitate Sebastian's series I came up with the patch below.
> The problem I have is, I do not really want the clkc to map the
> registers. They are in the SLCR and the SLCR driver is doing it, hence
> we should work with what that driver provides - which ideally would be
> based on regmap and syscon, but we're not there yet. Hence I somehow
> need to pass the SLCR pointer to the clkc. To avoid accessing the global
> pointer directly I kept the zynq_clock_init() routine which is called
> from zynq_slcr_init().

For this patch set I'd be fine with the proposal below. For the short
run, you could consider to hide register accesses to slcr by providing
zynq_slcr_readl/writel instead of passing just the base address.

But again, that will require either custom .init_time or .init_irq
to set up slcr before clocks.

> That is the best I could come up with quickly and w/o investing a lot of
> time to figure out the regmap and syscon stuff, which seems to be handled
> by Steffen already, anyway.
> It is essentially a stripped down version of Sebastian's proposal.

If there are no general objections, I take that one for the real patch
set.

Sebastian