Re: [PATCH 3/6] arm64: tegra: Add Tegra210 support

On Fri, May 22, 2015 at 02:38:00PM +0100, Catalin Marinas wrote:

On Tue, May 19, 2015 at 05:17:42PM +0200, Thierry Reding wrote:

quoted

On Tue, May 19, 2015 at 03:52:11PM +0100, Catalin Marinas wrote:

quoted

On Fri, May 15, 2015 at 12:19:16PM +0200, Thierry Reding wrote:

quoted

On Wed, May 13, 2015 at 06:11:15PM +0100, Catalin Marinas wrote:

quoted

On Wed, May 13, 2015 at 04:57:42PM +0200, Thierry Reding wrote:

quoted

From: Thierry Reding <redacted>

NVIDIA Tegra210 (also known as Tegra X1) has four Cortex-A57 and four
Cortex-A53 CPUs. Compared to Tegra124 and Tegra132 it comes with a 256-
core Maxwell GPU. It supports processing videos of up to 4K resolutions
at 60 fps (H.265, VP9, H.264).

Signed-off-by: Thierry Reding <redacted>
---
 arch/arm64/Kconfig                       |   9 +
 arch/arm64/boot/dts/nvidia/tegra210.dtsi | 998 +++++++++++++++++++++++++++++++
 2 files changed, 1007 insertions(+)
 create mode 100644 arch/arm64/boot/dts/nvidia/tegra210.dtsi

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 7796af4b1d6f..bfdf064ada66 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig

@@ -225,6 +225,15 @@ config ARCH_TEGRA_132_SOC
 	  but contains an NVIDIA Denver CPU complex in place of
 	  Tegra124's "4+1" Cortex-A15 CPU complex.
 
+config ARCH_TEGRA_210_SOC
+	bool "NVIDIA Tegra210 SoC"
+	depends on ARCH_TEGRA
+	select PINCTRL_TEGRA210
+	select USB_ULPI if USB_PHY
+	select USB_ULPI_VIEWPORT if USB_PHY
+	help
+	  Enable support for the NVIDIA Tegra210 SoC.
+

The previous ARCH_TEGRA_132_SOC escaped me. Do we need all these
ARCH_TEGRA_*_SOC entries? Can we not have per-driver Kconfig options?
For example, ARCH_TEGRA_132_SOC seems to be only used in
drivers/clk/tegra, a specific Kconfig entry in there would suffice.

There are actually a couple of other places where this will be used in
subsequent patches (e.g. memory controller driver). The idea behind
having these is that each one of them is used to enable the essentials
out of the box, so that people don't have to go and enable a bunch of
driver-specific Kconfig options just to get a kernel configuration that
can actually boot.

We debated whether to have ARCH_* options at all on arm64 and we settled
for the middle ground - only add them for SoC families, not individual
SoCs. As for the kernel configuration that actually boots, we want the
arm64 defconfig to include all the supported SoCs and drivers (though
longer term I'd like to see more drivers built as modules by default).

quoted

This is also useful for integrators since they can simply omit all SoC
generations that they're not interested in. Having a per-SoC option
provides an easy way of doing so.

The integrators could just select a SoC family and trim down unwanted
options, I don't think they rely on the kernel defconfig for a final
product. If this becomes an issue, I would rather have per-SoC
defconfigs than lots of Kconfig entries.

I understand the desire to start with a clean plate on a new
architecture, but Tegra has worked like this for the past 5 years and
it's worked out really well for us. So I'm reluctant to introduce these
inconsistencies merely because 64-bit ARM now lives in a different
directory.

That's a good time to start cleaning this up, especially since we
mandate single image from the beginning.

quoted

Are you concerned about arch/arm64/Kconfig growing wild? If so we could
easily move these configuration options outside to something like
drivers/soc/tegra/Kconfig. While at it, we could move existing options
from arch/arm/mach-tegra over to that as well.

I think the whole ARCH_TEGRA_*_SOC approach is wrong. You don't bother
introducing Kconfig entries for individual drivers but instead add an
obj-*_SOC in various Makefiles under drivers/. If we ever get to a point
where we can build (part of) the SoCs as modules (and I'm not talking
about kernels shipped with end products but those distro kernel
targeting development boards), we have to go back and add such Kconfig
entries for specific drivers.

We can cross that bridge when we get to it. Currently we only use them
in the fashion you describe for essential drivers that can't be built as
modules. If that ever changes we can easily add the Kconfig entry. This
isn't as much churn as you make it out to be, after all if the Kconfig
entry already existed we'd still have to change it from bool to tristate
to make them build as modules.

Another situation in which we use these Kconfig symbols is for drivers
that have an own Kconfig entry but that support more than one SoC
generation so that we automatically strip support for SoCs for which
support isn't selected. This is especially useful to avoid building code
that will only ever work on 32-bit Tegra on 64-bit builds.

So if you do it properly from the beginning with per-driver Kconfig
entry, I no longer see the point of an ARCH_TEGRA_*_SOC option. It may
be more convenient for people building end products but they should know
better what components are required, defconfig is rarely meant for them.

Your proposal also means that we'll have to go and add separate Kconfig
entries for each SoC generation that a driver supports so that the SoC
specific bits can be disabled.

So in order to get rid of one per-SoC Kconfig entry you'd prefer us to
add per-SoC options per driver? Don't you think that's going to make
things worse rather than improve them?

There are currently three drivers (off the top of my head) where we do
this and each of them supports five existing generations plus the new
Tegra210. So rather than having six symbols to select SoC generation
support we'd be needing 18, and that's only counting existing drivers.

I'm looking forward to a patch removing ARCH_TEGRA_132_SOC as well (you
can keep the 32-bit arm options if you'd like).

Given the above, will you reconsider? I already proposed that we move
these under drivers/soc/tegra (though I'd like to hear Arnd's opinion on
that first), that way you wouldn't have to concern yourself with them.

Thierry