[PATCHv4 11/15] clk: ti: clockdomain: add clock provider support to clockdomains

[Tony Lindgren]

* Michael Turquette <mturquette at baylibre.com> [161202 15:52]:
> Quoting Tony Lindgren (2016-12-02 15:12:40)
> > * Michael Turquette <mturquette at baylibre.com> [161202 14:34]:
> > > Quoting Tony Lindgren (2016-10-28 16:54:48)
> > > > * Stephen Boyd <sboyd at codeaurora.org> [161028 16:37]:
> > > > > On 10/28, Tony Lindgren wrote:
> > > > > > * Tero Kristo <t-kristo at ti.com> [161028 00:43]:
> > > > > > > On 28/10/16 03:50, Stephen Boyd wrote:
> > > > > > > > I suppose a PRCM is
> > > > > > > > like an MFD that has clocks and resets under it? On other
> > > > > > > > platforms we've combined that all into one node and just had
> > > > > > > > #clock-cells and #reset-cells in that node. Is there any reason
> > > > > > > > we can't do that here?
> > > > > > > 
> > > > > > > For OMAPs, there are typically multiple instances of the PRCM around; OMAP4
> > > > > > > for example has:
> > > > > > > 
> > > > > > > cm1 @ 0x4a004000 (clocks + clockdomains)
> > > > > > > cm2 @ 0x4a008000 (clocks + clockdomains)
> > > > > > > prm @ 0x4a306000 (few clocks + resets + power state handling)
> > > > > > > scrm @ 0x4a30a000 (few external clocks + plenty of misc stuff)
> > > > > > > 
> > > > > > > These instances are also under different power/voltage domains which means
> > > > > > > their PM behavior is different.
> > > > > > > 
> > > > > > > The idea behind having a clockdomain as a provider was mostly to have the
> > > > > > > topology visible : prcm-instance -> clockdomain -> clocks
> > > > > > 
> > > > > > Yeah that's needed to get the interconnect hierarchy right for
> > > > > > genpd :)
> > > > > > 
> > > > > > > ... but basically I think it would be possible to drop the clockdomain
> > > > > > > representation and just mark the prcm-instance as a clock provider. Tony,
> > > > > > > any thoughts on that?
> > > > > > 
> > > > > > No let's not drop the clockdomains as those will be needed when we
> > > > > > move things into proper hierarchy within the interconnect instances.
> > > > > > This will then help with getting things right with genpd.
> > > > > > 
> > > > > > In the long run we just want to specify clockdomain and the offset of
> > > > > > the clock instance within the clockdomain in the dts files.
> > > > > > 
> > > > > 
> > > > > Sorry, I have very little idea how OMAP hardware works. Do you
> > > > > mean that you will have different nodes for each clockdomain so
> > > > > that genpd can map 1:1 to the node in dts? But in hardware
> > > > > there's a prcm that allows us to control many clock domains
> > > > > through register read/writes? How is the interconnect involved?
> > > > 
> > > > There are multiple clockdomains, at least one for each interconnect
> > > > instance. Once a clockdomain is idle, the related interconnect can
> > > > idle too. So yeah genpd pretty much maps 1:1 with the clockdomains.
> > > > 
> > > > There's more info in for example omap4 TRM section "3.4.1 Device
> > > > Power-Management Layout" that shows the voltage/power/clock domains.
> > > > The interconnect instances are mostly named there too looking at
> > > > the L4/L3 naming.
> > > 
> > > I'm confused on two points:
> > > 
> > > 1) why are the clkdm's acting as clock providers? I've always hated the
> > > name "clock domain" since those bits are for managing module state, not
> > > clock state. The PRM, CM1 and CM2 provide the clocks, not the
> > > clockdomains.
> > 
> > The clock domains have multiple clock inputs that are routed to multiple
> > child clocks. So it is a clock :)
> > 
> > See for example omap4430 TRM "3.6.4 CD_WKUP Clock Domain" on page
> > 393 in my revision here.
> > 
> > On that page "Figure 3-48" shows CD_WKUP with the four input clocks.
> > And then "Table 3-84. CD_WKUP Control and Status Parameters" shows
> > the CD_WKUP clock domain specific registers. These registers show
> > the status, I think they are all read-only registers. Then CD_WKUP
> > has multiple child clocks with configurable registers.
> > 
> > From hardware register point of view, each clock domain has:
> > 
> > - Read-only clockdomain status registers in the beginning of
> >   the address space
> > 
> > - Multiple similar clock instances register instances each
> >   mapping to a specific interconnect target module
> > 
> > These are documented in "3.11.16.1 WKUP_CM Register Summary".
> 
> Oh, this is because you are treating the MODULEMODE bits like gate
> clocks. I never really figured out if this was the best way to model
> those bits since they do more than control a line toggling at a rate.
> For instance this bit will affect the master/slave IDLE protocol between
> the module and the PRCM.

Yes seems like there is some negotiation going on there with the
target module. But from practical point of view the CLKCTRL
register is the gate for a module functional clock.

> > From hardware point of view, we ideally want to map interconnect
> > target modules to the clock instance offset from the clock domain
> > for that interconnect segment. For example gptimer1 clocks would
> > be just:
> > 
> > clocks = <&cd_wkup 0x40>;
> > 
> > > 2) why aren't the clock domains modeled as genpds with their associated
> > > devices attached to them? Note that it is possible to "nest" genpd
> > > objects. This would also allow for the "Clockdomain Dependency"
> > > relationships to be properly modeled (see section 3.1.1.1.7 Clock Domain
> > > Dependency in the OMAP4 TRM).
> > 
> > Clock domains only route clocks to child clocks. Power domains
> > are different registers. The power domains map roughly to
> > interconnect instances, there we have registers to disable the
> > whole interconnect when idle.
> 
> I'm not talking about power islands at all, but the genpd object in
> Linux. For instance, if we treat each clock domain like a clock
> provider, how could the functional dependency between clkdm_A and
> clkdm_B be asserted?

To me it seems that some output of a clockdomain is just a input
of another clockdomain? So it's just the usual parent child
relationship once we treat a clockdomain just as a clock. Tero
probably has some input here.

> There is certainly no API for that in the clock framework, but for genpd
> your runtime_pm_get() callback for clkdm_A could call runtime_pm_get
> against clkdm_B, which would satisfy the requirement. See section
> 3.1.1.1.7 Clock Domain Dependency in the OMAP4 TRM, version AB.

To me it seems the API is just clk_get() :) Do you have some
specific example we can use to check? My guess is that the
TRM "Clock Domain Dependency" is just the usual parent child
relationship between clocks that are the clockdomains..

If there is something more magical there certainly that should
be considered though.

Regards,

Tony