Re: [PATCH] clk: Register clkdev after setup of fixed-rate and fixed-factor clocks

[Xiaolong Zhang]

On ä, 11æ 23, 2016 at 04:38:33äå -0800, Stephen Boyd wrote:
> On 11/12, Xiaolong Zhang wrote:
> > On ä, 10æ 25, 2016 at 08:40:08äå +0000, Stephen Boyd wrote:
> > > On 10/22, Xiaolong Zhang wrote:
> > > > On å, 10æ 20, 2016 at 04:01:03äå -0700, Stephen Boyd wrote:
> > > > > On 10/11, Orson Zhai wrote:
> > > > > > From: Xiaolong Zhang <xiaolong.zhang@xxxxxxxxxxxxxx>
> > > > > > 
> > > > > > When common kernel setups fixed clock, of_clk_provider will be registerred.
> > > > > > But there is no clkdev being registerred at the same time. This will make
> > > > > > it difficult to get the clock by using clk_get(NULL, con_id).
> > > > > > 
> > > > > > Add clkdev register for fixed-rate and fixed-factor clock and ignore
> > > > > > the error if any.
> > > > > > 
> > > > > > Signed-off-by: Xiaolong Zhang <xiaolong.zhang@xxxxxxxxxxxxxx>
> > > > > > Signed-off-by: Orson Zhai <orson.zhai@xxxxxxxxxxxxxx>
> > > > > > ---
> > > > > 
> > > > > Why are we using clkdev lookups for clks populated from DT?
> > > > > Shouldn't we be able to point to them from the consumers that
> > > > > would also be in DT? I'm a little lost.
> > > > > 
> > > > The clk_get interface allows the first argument as NULL. We just assure
> > > > consumers can get the clock from DT or by clock name.
> > > 
> > > Ok. The first argument to clk_get() really shouldn't be NULL. It
> > > should be the device pointer for the device that is associated
> > > with the clk you want to get. The clock name (second argument to
> > > clk_get()) should be device specific and not some globally unique
> > > identifier. It seems that you're using the clk_get() API
> > > incorrectly.
> > >
> > Sorry for late reply.
> > There are two paths in clk_get:
> > a) dev is not NULL, the calling procedure is as following
> > clk_get(dev, con_id)
> > --->of_clk_get_by_name(dev->of_node, con_id)
> > 
> > b) dev is NULL, the calling procedure is as following
> > clk_dev(NULL, con_id)
> > --->clk_get_sys(NULL, con_id)
> >     --->clk_find(NULL, con_id)
> > 
> > I just cann't understand why you say the first argument shouldn't
> > be NULL. Is there other consideration in CCF?
> > 
> 
> Passing NULL as the first argument is allowed, but that's mostly
> an artifact of the clk_get() API. If you don't have a device, you
> should be calling clk_get_sys() and then a clkdev lookup should
> have been added with some appropriate dev_id string that is used
> in the clk_get_sys() call. At least this is my understanding of
> the clkdev APIs, but Russell is the authority here.
>

I understand what you mean. If a driver uses the device tree, the
consumers and providers have the fixed relationship. But I think
clk_get() has a purpose is compatible with the non-DT driver.

> We're really off track now though. Can you please point to some
> code that needs this change? If we're using DT then we should be
> able to use the of_clk_*() path to find the clk.
>

Actually, the requirement is raised by our GPU driver. In the
early stage of the GPU DT driver, the GPU driver use the
clk_get(NULL, con_id) to get the clock instance for compatible
with non-DT GPU driver. The new driver have used the of_clk_get()
instead of the clk_get. And we reserved the modification in clock.

Xiaolong Zhang
Thanks
> -- 
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> a Linux Foundation Collaborative Project