[PATCH v4 4/7] cpufreq: add clk-reg cpufreq driver

[Richard Zhao]

On 3 January 2012 21:47, Russell King - ARM Linux
<linux at arm.linux.org.uk> wrote:
> On Tue, Jan 03, 2012 at 09:25:30PM +0800, Richard Zhao wrote:
>> Hi Russel,
>>
>> On 3 January 2012 17:06, Russell King - ARM Linux
>> <linux at arm.linux.org.uk> wrote:
>> > On Mon, Dec 26, 2011 at 09:44:52PM +0800, Richard Zhao wrote:
>> >> On Mon, Dec 26, 2011 at 11:10:30AM +0000, Mark Brown wrote:
>> >> > The *call* is there in the regulator subsystem, it's just that none of
>> >> > the drivers back it up with an actual implementation yet.  Which turns
>> >> > out to be a good thing as cpufreq can't currently understand variable
>> >> > latencies and the governors don't deal well with non-trivial latencies
>> >> > anyway.
>> >> but clk API don't have such calls. and many SoCs only adjust clk frequencies, using
>> >> one single voltage.
>> >
>> > That's because it's often not known - especially in the case of PLLs,
>> > data sheets don't tend to specify how long it takes for the PLL to relock
>> > after a requested change.  If it's important that the PLL be locked,
>> > there will be a bit to poll (or they'll cause the CPU itself to stall
>> > while the PLL is not locked.)
>> >
>> > So, for these kinds of situations, how do you suggest that the clk API
>> > provides this information?
>> In latest v6 version, I get clk transition latency from dt property, and get
>> regulator transition latency from regulator API.
>> Could you please help review other arm common changes in v6 version?
>
> You didn't get my point: how do you specify a clock transition latency
> for a clock with a PLL when the data sheets don't tell you what that is,
> and they instead give you a bit to poll?
clk-trans-latency is not a MUST property. If the datasheet don't provide it or
you don't want to measure, you may not specify the property. In such case
cpufreq governor ondemand and conservative will not work, because they
can not determine the sample rate without the latency value.
>
> Do you:
>
> (a) make up some number and hope that it's representative
No, I always get it from dt.

> (b) not specify any transition latency
It's allowed.

> (c) think about the problem _now_ and define what it means for a clock
>    without a transition latency.
It's ondemand/conservative governors that strongly need the latency value.
If user don't want to use it, it's fine.

Thanks
Richard