[PATCH 2/2] clk: initial clock driver for TWL6030

[Mike Turquette]

Quoting Stefan Assmann (2014-07-31 05:04:29)
> On 31.07.2014 13:05, Mark Brown wrote:
> > On Thu, Jul 31, 2014 at 11:56:15AM +0200, Stefan Assmann wrote:
> >> On 30.07.2014 19:50, Mark Brown wrote:
> >>> On Wed, Jul 30, 2014 at 04:02:29PM +0200, Stefan Assmann wrote:
> > 
> >>>> +static int twl6030_clk32kg_is_prepared(struct clk_hw *hw)
> >>>> +{
> >>>> +  struct twl6030_desc *desc = to_twl6030_desc(hw);
> >>>> +
> >>>> +  return desc->enabled;
> >>>> +}
> > 
> >>> Why not just check the register map - can't the register be cached?  If
> >>> that's not possible a comment would be good.
> > 
> >> I just took atl_clk_is_enabled() as template. If you say it's better
> >> to read the value, that can be arranged.
> > 
> > It might be worth doing this if you have to go to hardware to check the
> > status, if you can read a cache then just using the register is less
> > error prone.
> 
> Ok.
> 
> > 
> >>>> +  else
> >>>> +          clk_prepare(clk);
> > 
> >>> Why is the clock driver defaulting to enabling the clock, and if it
> >>> needs to shouldn't it be doing a prepere_enable() even if the enable
> >>> happens not to do anything to the hardware?  Otherwise child clocks
> >>> might get confused.
> > 
> >> Mike advised me to convert the functions from enable/disable to
> >> prepare/unprepare because i2c transactions may sleep. That's what I did.
> >> The code no longer enables the clock and just prepares it. So IIUC the
> >> call to clk_prepare() should be fine.
> > 
> > That's not going to help consumers of the clock, you do need to move the
> > operations to prepare() but users shouldn't need to know what happens in
> > prepare() and what happens in enable().
> > 
> > You've also not addressed the comment about defaulting to enabling the
> > clock in the first place.
> 
> Maybe I misinterpreted your previous comment, sorry. So if I got it right
> this time you're saying that the prepare/enable, disable/unprepare should
> be seen as a single step. If that's the case then using prepare_enable()
> should be used indeed.

It's not so much that they are single step. The point is that the way we
ungate a clock signal using the Linux clk.h api is by calling
clk_enable().

If you read the code for clk_enable you'll see that it depends on the
clock already having been prepared via clk_prepare(). The fact that your
particular implementation of this clock driver actually ungates the
physical signal via clk_prepare is irrelevant from the api's
perspective.  clk_enable is how we turn gateable clocks on, even if the
hardware really gets affected by the prior call to clk_prepare.

It would be a real mess if drivers knew the details of their underlying
clock hardware and only called clk_prepare and never clk_enable to
gate/ungate their clocks.

Regards,
Mike

> 
> Tero suggested to look into making this a generic i2c clock driver. I'll
> look into that and report back.
> 
>   Stefan