* Tero Kristo [off-list ref] [170123 10:30]:

On 23/01/17 18:28, Tony Lindgren wrote:

quoted

* Tero Kristo [off-list ref] [170123 06:45]:

quoted

On 18/01/17 00:53, Tony Lindgren wrote:

quoted

Texas Instruments omap variant SoCs starting with omap4 have a clkctrl
clock controller instance for each interconnect target module. The clkctrl
controls functional and interface clocks for the module.

The clkctrl clocks are currently handled by arch/arm/mach-omap2 hwmod code.
With this binding and a related clock device driver we can start moving the
clkctrl clock handling to live in drivers/clk/ti.

Note that this binding allows keeping the clockdomain related parts out of
drivers/clock. The CLKCTCTRL and DYNAMICDEP registers can be handled by
a separate driver in drivers/soc/ti and genpd. If the clockdomain driver
needs to know it's clocks, we can just set the the clkctrl device
instances to be children of the related clockdomain device.

Each clkctrl clock can have multiple optional gate clocks, and multiple
optional mux clocks. To represent this in device tree, it seems that
it is best done using four clock cells #clock-cells = <4> property.

The reasons for using #clock-cells = <4> are:

1. We need to specify the clkctrl offset from the instance base. Otherwise
   we end up with a large number of device tree nodes that need to be
   patched when new clocks are discovered in a clkctrl clock with minor
   hardware revision changes for example

2. On omap5 CM_L3INIT_USB_HOST_HS_CLKCTRL has ten OPTFCLKEN bits. So we
   need to use a separate cell for optional gate clocks to avoid address
   space conflicts

3. Some clkctrl instances can also also optional mux clocks. To address
   them properly we need also a separate cell for the optional mux
   clock index

4. The modulemode clock needs a flag passed to it for hardware or
   software controlled mode

Hi Tony,

I think #clock-cells = <4> is too much. I believe we only need 2:

- one entry for clkctrl offset
- one entry for clock offset within the clkctrl entry (0 = module clock, 8+
= opt-clocks / mux clocks / dividers)

OK the less #clock-cells the better as long as it's enough :)

quoted

Fields 2 / 3 in your proposal are mutually exclusive, if either field is
non-zero, the other one must be zero. And, the opt clocks / mux / divs
always have different values for these.

OK. Just to confirm the assumptions then:

1. The optional mux clock the consumer needs to select the right
   source clock with with clk_set_parent()

Yes. And for this you need to fetch a clock handle via some mechanism
(of_clk_get, clk_get...) Clock consumers can't directly use parent IDs.

quoted

2. The optional divider clock rate must be set by the consumer
   using clk_set_rate()

Yes again.

quoted

And in that case we again don't need to define any artificial
clock indexes, which is good if new clocks are discovered between
various SoC revisions.

quoted

Field 4 is kind of redundant also, as the module clock must be registered at
the clkctrl probe time, it is too late for the clock consumer to provide the
proper setting for the clock during its own probe. It seems I need to add
static data to driver which basically has this information in place already.

OK yeah good point, the "clocks" is a consumer property.

So in that case we must also assume that if any clock consumer needs
to change between HWSUP or SWSUP, it needs to be done with some yet
to be determined API. We have not needed that so far AFAIK though.

If there are no issues with the above, I'm naturally fine using the
#clock-cells = <2> :)

Yeah, clock-cells = <2>; seems to work just fine in the WIP codebase I have.

OK thanks for confirming, will post v3 of the binding.

Regards,

Tony
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html