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

[SÃren Brinkmann]

Hi Sebastian, Steffen,

On Fri, Aug 23, 2013 at 11:30:18AM +0200, Sebastian Hesselbarth wrote:
> On 08/23/13 02:59, SÃren Brinkmann wrote:
> >On Thu, Aug 22, 2013 at 05:26:47PM -0700, SÃren 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?

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.

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().

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.

	SÃren

-----8<--------------------8<-------------------8<-------------------