Re: [PATCH v2 4/4] clk: dt: Introduce always-on clock domain documentation

[Lee Jones <lee.jones@linaro.org>]

On Thu, 19 Feb 2015, Geert Uytterhoeven wrote:
> On Thu, Feb 19, 2015 at 11:43 AM, Lee Jones <lee.jones@linaro.org> wrote:
> > On Thu, 19 Feb 2015, Geert Uytterhoeven wrote:
> >> On Thu, Feb 19, 2015 at 11:28 AM, Lee Jones <lee.jones@linaro.org> wrote:
> >> > On Thu, 19 Feb 2015, Geert Uytterhoeven wrote:
> >> >> On Thu, Feb 19, 2015 at 11:11 AM, Lee Jones <lee.jones@linaro.org> wrote:
> >> >> > On Thu, 19 Feb 2015, Geert Uytterhoeven wrote:
> >> >> >> On Thu, Feb 19, 2015 at 10:42 AM, Lee Jones <lee.jones@linaro.org> wrote:
> >> >> >> >> What kind of clocks are these? What do they control?
> >> >> >> >> Memory controllers? Bus controllers?
> >> >> >> >>
> >> >> >> >> They must control some device(s), so there should be one or more device
> >> >> >> >> nodes in DT that reference these clocks.
> >> >> >> >> As soon as that information is in DT, support can be added to Linux to
> >> >> >> >> make sure the "critical" clocks stay enabled, either through a real driver,
> >> >> >> >> or through platform code.
> >> >> >> >
> >> >> >> > Some do, some don't.  For instance, we have one clock which controls
> >> >> >> > SPI and I2C that must not be turned off.  We discovered this then when
> >> >> >> > a suspend was attempted and the board refused to resume.  This clock
> >> >> >> > also runs one of the critical interconnects that runs from the a9.  It
> >> >> >> > would be wrong to remove the clk_disable() attempt from the SPI/I2C
> >> >> >> > drivers because the same IP on another board might be controlled by a
> >> >> >> > different clock which is able to be gated.
> >> >> >> >
> >> >> >> > There are also clocks which control other interconnects that are not
> >> >> >> > connected to any device drivers.  If we fail to take references for
> >> >> >> > them before clk_disable_unused() is called, again the board hangs.  We
> >> >> >> > even lose JTAG support.
> >> >> >>
> >> >> >> Interconnects are buses. Can't you represent those buses in the DT
> >> >> >> hierarchy, and give them clocks properties?
> >> >> >
> >> >> > So instead of this nice succinct, simple, cover all bases
> >> >> > (interconnects was just an example, there are bound to be others),
> >> >> > generic framework, you are suggesting to write drivers for devices
> >> >> > which other than "don't turn my clocks off", Linux can't actually see
> >> >> > or control?
> >> >>
> >> >> DT describes the hardware, not behavior.
> >> >
> >> > Okay so ...
> >> >
> >> > /*
> >> >  * ICNs are not visible/controllable in Linux, but references to their
> >> >  * clocks must be obtained and retained or the platform will become
> >> >  * irrecoverably unresponsive.
> >> >  */
> >> > interconnects@0 {
> >> >        compatible = "always-on-clk-domain";
> >>
> >> st,...flexgen...
> >>
> >> >        clocks = <&clk_s_c0_flexgen CLK_ICN_SBC>,
> >> >                 <&clk_s_c0_flexgen CLK_ICN_LMI>,
> >> >                 <&clk_s_c0_flexgen CLK_ICN_CPU>,
> >> >                 <&clk_s_c0_flexgen CLK_TX_ICN_DMU>;
> >> > };
> >>
> >> And then you can have platform code that binds against st,...flexgen...,
> >> and enables all referenced clocks.
> >
> > Flexgen isn't a device, it's a clk source.  a) writing a device driver
> 
> Sorry, I'm not familiar with ST nomenclature.
> So that should become the name of the interconnect/bus.

You're still talking about writing function-less drivers for multiple
pieces of h/w.  We would have a few for ST alone, then multiply that
by the number of silicon vendors with similar issues -- which is
likely to be most of them.

I can understand Rob's "DT has to match h/w" point, but to insist we
write lots of empty drivers just to stop some clocks from being gated
is barking mad.

> > for a clk source seems wrong b) what if on another platform a
> > different clock source supplied the clock?  Write another driver?  And
> > what if the ICNs are connected to different clock sources?  More
> > drivers?  c) all of these drivers will only do one thing -- pull a
> > reference and keep hold of it.  You want 50 drivers (across all
> > platforms) doing only that?  Or, more sanely, do you want this one
> > generic framework driver doing that?
> >
> >> Alternatively, if you have power domains, you can add a reference to
> >> the power domain, and let the power domain driver handle it.
> >
> > I'm not sure what a power domain driver will do?  We need a driver to
> > _not_ give up references, that is all. :)
> 
> A power domain driver can do anything it wants.
> That includes enabling your interconnect clocks, and keeping them enabled.
> 
> Now, if flexgen is the name of the clock source, then all devices connected
> to flexgen are part of the flexgen clock domain, which can be represented
> in Linux using the generic PM domain. Do all devices connected to flexgen
> need to be handled similarly w.r.t. clocks? If yes, the genpd may be the place
> to implement that.

Most of the FLEXGEN clocks are fully gateable.  It's only a select few
which are critical to the running of the system.

-- 
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org │ Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog