Re: [RFC v2 4/9] of: add clock providers

[Grant Likely <grant.likely@secretlab.ca>]

On Tue, Jan 17, 2012 at 4:37 PM, Turquette, Mike <mturquette@ti.com> wrote:
> On Tue, Jan 17, 2012 at 2:47 PM, Grant Likely <grant.likely@secretlab.ca> wrote:
>> On Tue, Jan 17, 2012 at 1:44 PM, Stephen Warren <swarren@nvidia.com> wrote:
>>> In other words, does the UART driver need to do something like:
>>>
>>> clk_reg = clk_get(dev, "register");
>>> clk_parent = of_clk_get_by_name(np, "register);
>>> clk_set_parent(clk_reg, clk_parent);
>>>
>>> Or will that all happen transparently within just the of_clk_get_by_name
>>> call?
>>>
>>> (I suppose this question makes slightly more sense for the PLL itself,
>>> since both the upstream and downstream clocks are represented in the PLL
>>> node, whereas the UART's node only represents the clock consumer side,
>>> so the above code isn't really possible automatically).
>>
>> The intent is that device only interacts with the leaf device.  If the
>> clocks are arranged into a hierarchy, then the clock driver is
>> responsible for any interactions with the parent clock.  Requiring the
>> driver to manipulate parent clocks directly defeats the purpose of
>> having a clock abstraction.
>
> I don't think that we can get rid of all instances of drivers knowing
> a bit about hierarchy.  I think we can get rid of most, but there are
> cases where it is valid for a driver to know some of the details.
> More on that below.
>
>>> Somewhat related to this: How does dynamic reparenting interact with
>>> the DT clock binding; is the DT just the default/initial clock setup,
>>> and anything beyond that needs a custom binding and code in the consumer?
>>
>> As far as the clock binding goes, it only describes provider/consumer
>> relationships.  The fact that such relationships may resolve to a
>> hierarchy is beyond what the binding describes.  If a clock has
>> multiple possible parents, then that specific clock binding should
>> document how the multiple parent clocks are described and the clock
>> driver is responsible for implementing the correct behaviour.
>
> It also deserves to be said that the DT data says nothing about which
> of the possible parents _should_ be the input to a mux clock.  It's
> pretty common to want to make changes to hierarchy after taking a
> device out of reset, since the reset values for a clock management IP
> might be pretty conservative.  So someone, somewhere must know some
> details about hierarchy and set things up correctly.  Maybe a "clock
> driver" can do this, but for specific IPs such as the audio example
> below it makes sense for that driver to have the knowledge.
>
>> Similarly, the DT clock binding provides no generic mechanism for
>> walking up the clock tree.  That behaviour must also be implemented by
>> each specific clock driver.
>>
>>> I'm thinking of say a system with 1 I2S controller, and both an internal
>>> and external I2S clock source, where perhaps the internal source needs
>>> to be used to capture from an I2S interface on one set of pins (e.g.
>>> analog mic) but the other clock source needs to be used to capture from
>>> I2S on another set of pins (e.g. digital baseband unit connection).
>>> (This example is theoretical, but I'm sure there are other dynamic clock
>>> cases in practice).
>>
>> That is a reasonable example.  In this case, the i2c controller would
>> include both in its clocks property, and the binding would document
>> when and why each clock source is used.
>
> I'm confused on this point.  How does the binding "document when and
> why each clock is used"?  In the case where this I2S controller
> expects to dynamically switch roles at run-time (analog mic versus
> baseband) then clk_set_parent must still be invoked by the driver.  To
> be clear I'm imagining the above example like:
>
> i_I2S   e_I2S
>   \       /
>    I2S_mux
>        |
> I2S controller IP

>From the description, I assumed that the i2s_mux was part of the i2s
controller driver which would select the appropriate clock based on
the configured mode.  If it is better to implement it as a separate
clock driver, then yes the i2s controller must have knowledge of that
and call clk_set_parent appropriately.  Regardless, some driver needs
to explicitly understand the relationship between pinmux and clock
source.

g.