Hi Liu,
On Fri, Aug 12, 2022 at 10:13 AM Liu Ying <victor.liu@xxxxxxx> wrote:
> On Thu, 2022-08-11 at 14:34 +0200, Geert Uytterhoeven wrote:
> > On Thu, Aug 4, 2022 at 8:10 AM Liu Ying <victor.liu@xxxxxxx> wrote:
> > > Simple Power-Managed bus controller may need functional clock(s)
> > > to be enabled before child devices connected to the bus can be
> > > accessed. Get the clock(s) as a bulk and enable/disable the
> > > clock(s) when the bus is being power managed.
> > >
> > > One example is that Freescale i.MX8qxp pixel link MSI bus
> > > controller
> > > needs MSI clock and AHB clock to be enabled before accessing child
> > > devices.
> > >
> > > Signed-off-by: Liu Ying <victor.liu@xxxxxxx>
> >
> > Thanks for your patch!
>
> Thanks for the review.
>
> >
> > > --- a/drivers/bus/simple-pm-bus.c
> > > +++ b/drivers/bus/simple-pm-bus.c
> > > @@ -8,11 +8,17 @@
> > > * for more details.
> > > */
> > >
> > > +#include <linux/clk.h>
> > > #include <linux/module.h>
> > > #include <linux/of_platform.h>
> > > #include <linux/platform_device.h>
> > > #include <linux/pm_runtime.h>
> > >
> > > +struct simple_pm_bus {
> > > + struct clk_bulk_data *clks;
> > > + int num_clks;
> > > +};
> > > +
> > > static const struct of_device_id simple_pm_bus_child_matches[] = {
> > > { .compatible = "simple-mfd", },
> > > {}
> > > @@ -24,6 +30,7 @@ static int simple_pm_bus_probe(struct
> > > platform_device *pdev)
> > > const struct of_dev_auxdata *lookup =
> > > dev_get_platdata(dev);
> > > struct device_node *np = dev->of_node;
> > > const struct of_device_id *match;
> > > + struct simple_pm_bus *bus;
> > >
> > > /*
> > > * Allow user to use driver_override to bind this driver to
> > > a
> > > @@ -49,6 +56,16 @@ static int simple_pm_bus_probe(struct
> > > platform_device *pdev)
> > > return -ENODEV;
> > > }
> > >
> > > + bus = devm_kzalloc(&pdev->dev, sizeof(*bus), GFP_KERNEL);
> > > + if (!bus)
> > > + return -ENOMEM;
> > > +
> > > + bus->num_clks = devm_clk_bulk_get_all(&pdev->dev, &bus-
> > > > clks);
> > >
> > > + if (bus->num_clks < 0)
> > > + return dev_err_probe(&pdev->dev, bus->num_clks,
> > > "failed to get clocks\n");
> > > +
> > > + dev_set_drvdata(&pdev->dev, bus);
> > > +
> > > dev_dbg(&pdev->dev, "%s\n", __func__);
> > >
> > > pm_runtime_enable(&pdev->dev);
> >
> > While I agree this patch has merits on its own[*], I am wondering
> > if you really need it for your use case.
> >
> > In "[PATCH v3 3/3] dt-bindings: bus: Add Freescale i.MX8qxp pixel
> > link MSI bus binding", I see your bus has both "clocks" and
> > "power-domains" properties. Perhaps your PM Domain can be a clock
> > domain, too (i.e. setting GENPD_FLAG_PM_CLK and providing
> > generic_pm_domain.{at,de}tach_dev() callbacks), thus handing clock
> > handling over to Runtime PM?
>
> It looks like most(if not all) PM domains can be clock domains with
> GENPD_FLAG_PM_CLK and generic_pm_domain.{at,de}tach_dev() callbacks
> set. So, technically, my PM domain(scu-pd.c) can be a clock domain with
> all those set and a special check for "simple-pm-bus" in the
> {at,de}tach_dev callbacks. But, I'm not sure if it is appropriate to
> do that. How do we determine clocks should be managed by PM domains or
> device drivers? Technically, both would work...
It depends on the hardware topology: is there really a clock domain
(i.e. lots of consumer modules are clocked by a single clock
controller and can be power-managed that way), or is it just a
coincidence that your bus has clocks.
E.g. drivers/clk/renesas/renesas-cpg-mssr.c:cpg_mssr_attach_dev()
looks for clocks from the right clock provider and of the right type.
> > [*] The simple-pm-bus DT bindings state:
> >
> > clocks: true
> > # Functional clocks
> > # Required if power-domains is absent, optional otherwise
> >
> > power-domains:
> > # Required if clocks is absent, optional otherwise
> > minItems: 1
> >
> > While "power-domains" (+ "clocks" in case of a clock domain) is
> > handled through Runtime PM, the current driver indeed does not handle
> > "clocks" in the absence of the "power-domains" property. It looks
>
> Right.
>
> > like all existing users that use "clocks" rely on the PM Domain being
> > a clock domain.
>
> "renesas,bsc" seems to be one of the users.
Yes it is. And it doesn't need a special driver, as it just relies
on Runtime PM controlling both the power area and the clocks through
the PM Domain.
> > With your patch, the clocks might be controlled twice: once
> > explicitly,
> > through clk_bulk_*(), and a second time implicitly, through Runtime
> > PM.
> > While this works fine to do clock enable counters, it looks
> > suboptimal
> > to me. This could be avoided by making the new explicit clock code
> > depend on the absence of the "power-domains" property, but that would
> > break users that have both a PM Domain which is not a clock domain,
> > and clocks. So we may have no other option.
>
> Hmm, I'm not sure if there are really users that have both a PM domain
> which is not a clock domain and clocks, given that a PM domain can sort
> of always be a clock domain by setting that GENPD flag and those
> callbacks. So, what do you suggest? Keep the patch as-is? Or, maybe,
> make my PM domain additionally be a clock domain?
It depends. Is "dc0_disp_ctrl_link_mst0_lpcg" a clock controller that
controls the clock inputs to multiple modules? Based on the name, it
seems to be used only for display-related clocks, while "pd" controls
power to various modules, not limited to display?
Hence you may want to keep your patch as-is.
Renesas R-Car SoCs have separate power area and clock controllers,
too, but they apply to most/all devices. Hence we moduled this as
a single PM Domain (also because "power-domains" (used to) support
only a single provider), through a close integration of the power
area and clock drivers.
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
