On Wed, Dec 04, 2019 at 03:21:38PM +0530, Viresh Kumar wrote:
> On 04-12-19, 10:33, Thierry Reding wrote:
> > Yeah, the code that registers this device is in drivers/base/cpu.c in
> > register_cpu(). It even retrieves the device tree node for the CPU from
> > device tree and stores it in cpu->dev.of_node, so we should be able to
> > just pass &cpu->dev to tegra_bpmp_get() in order to retrieve a reference
> > to the BPMP.
> > 
> > That said, I'm wondering if perhaps we could just add a compatible
> > string to the /cpus node for cases like this where we don't have an
> > actual device representing the CPU complex. There are a number of CPU
> > frequency drivers that register dummy devices just so that they have
> > something to bind a driver to.
> > 
> > If we allow the /cpus node to represent the CPU complex (if no other
> > "device" does that yet), we can add a compatible string and have the
> > cpufreq driver match on that.
> > 
> > Of course this would be slightly difficult to retrofit into existing
> > drivers because they'd need to remain backwards compatible with existing
> > device trees. But it would allow future drivers to do this a little more
> > elegantly. For some SoCs this may not matter, but especially once you
> > start depending on additional resources this would come in handy.
> > 
> > Adding Rob and the device tree mailing list for feedback on this idea.
> 
> Took some time to find this thread, but something around this was
> suggested by Rafael earlier.
> 
> https://lore.kernel.org/lkml/8139001.Q4eV8YG1Il@vostro.rjw.lan/

I gave this a try and came up with the following:

--- >8 ---
diff --git a/arch/arm64/boot/dts/nvidia/tegra194.dtsi b/arch/arm64/boot/dts/nvidia/tegra194.dtsi
index f4ede86e32b4..e4462f95f0b3 100644
--- a/arch/arm64/boot/dts/nvidia/tegra194.dtsi
+++ b/arch/arm64/boot/dts/nvidia/tegra194.dtsi
@@ -1764,6 +1764,9 @@ bpmp_thermal: thermal {
 	};
 
 	cpus {
+		compatible = "nvidia,tegra194-ccplex";
+		nvidia,bpmp = <&bpmp>;
+
 		#address-cells = <1>;
 		#size-cells = <0>;
 
--- >8 ---

Now I can do something rougly like this, although I have a more complete
patch locally that also gets rid of all the global variables because we
now actually have a struct platform_device that we can anchor everything
at:

--- >8 ---
static const struct of_device_id tegra194_cpufreq_of_match[] = {
	{ .compatible = "nvidia,tegra194-ccplex", },
	{ /* sentinel */ }
};
MODULE_DEVICE_TABLE(of, tegra194_cpufreq_of_match);

static struct platform_driver tegra194_ccplex_driver = {
	.driver = {
		.name = "tegra194-cpufreq",
		.of_match_table = tegra194_cpufreq_of_match,
	},
	.probe = tegra194_cpufreq_probe,
	.remove = tegra194_cpufreq_remove,
};
module_platform_driver(tegra194_ccplex_driver);
--- >8 ---

I don't think that's exactly what Rafael (Cc'ed) had in mind, since the
above thread seems to have mostly talked about binding a driver to each
individual CPU.

But this seems a lot better than having to instantiate a device from
scratch just so that a driver can bind to it and it allows additional
properties to be associated with the CCPLEX device.

Rob, any thoughts on this from a device tree point of view? The /cpus
bindings don't mention the compatible property, but there doesn't seem
to be anything in the bindings that would prohibit its use.

If we can agree on that, I can forward my local changes to Sumit for
inclusion or reference.

Thierry