[PATCH 00/44] outer cache changes

[gregory.clement@free-electrons.com (Gregory CLEMENT)]

Hi Russell,

On 17/03/2014 01:13, Russell King - ARM Linux wrote:
> This patch series is where I'm currently at with the total mess which
> is L2 cache support.  That's putting it mildly.  This isn't intended
> for anyone to pick up - this is still mostly in development.
> 
> I need help with the Aurora/Tauros3/Broadcom stuff, as that remains
> addicted to the old buggy code that we used to have (eg, if you build
> a v6+v7 kernel, all PL310 errata get disabled.  If CONFIG_CACHE_PL310
> is enabled and you actually have a L2C-220, the necessary waits will
> be omitted.  If you enable erratum 588369 and set the .set_debug
> method to NULL or have a PL310 later than R3P0, you're opening a
> race condition. etc.)
> 
> One of the potentially explosive issues is the utter crap and fscked
> state of DT implementation here, much of which /can't/ now be fixed
> without breaking existing DT based booting with existing files,
> because we have a mixture of DT properties combined with platform
> specific code setting various L2C configuration parameters (sometimes
> writing directly to the registers.)
> 
> For example, I've made a start at trying to prevent any further additions
> of platform code writing to the L2C-310's power control register by
> adding a couple of properties - but we can't delete the direct register
> write in iMX6 code without breaking existing DT descriptions.
> 
> The same goes for the mess which is the auxillary control register,
> or now the prefetch control register.  Or the erratum 752271 which
> iMX6 implements in their own SoC specific code.
> 
> Here's the *big* explosive question: how have we ended up in this mess?
> A bigger question: why was none of this caught at review time?  An even
> bigger question (which I believe is at the root of much of the pain that
> people feel with DT): why aren't device properties being thought out
> better - by that I mean, why isn't the question "are the properties you
> are proposing for device X sufficient to describe all it's configurable
> parameters, and if not, what are you omitting?" being asked when new
> device bindings are being proposed.
> 
> You can see the extent of the L2C crap from the diffstat below - every
> arch/arm/mach-* file touched here is a symptom of failing to adequately
> describe the properties that platforms need to specify.  It's a failing
> of DT process.  It's a failing of review of SoC code.
> 

I had a quick look on this series and it is a really good improvement of
the code. I didn't notice anything suspicious regarding the Aurora L2
cache controller.

Then I have applied this series on 3.14-rc8 using our mvebu_defconfig. And
made some performance tests using cachebench. So I tested it on:

- an Armada XP based board which use the Aurora cache controller with the
aurora-system-cache flavor.

- an Armada 370 based board which use the Aurora cache controller with the
aurora-outer-cache.

I didn't notice any regression in both case.

So at least you can add my

Tested-by: Gregory CLEMENT <gregory.clement@free-electrons.com>


Thanks for your work,

Gregory


> Anyway, here's the diffstat so far.  Not Cc'ing anyone this time around
> (the Cc list is huge), so it's all on LAKML.
> 
>  Documentation/devicetree/bindings/arm/l2cc.txt |    2 +
>  arch/arm/Kconfig                               |   51 -
>  arch/arm/boot/dts/imx6qdl.dtsi                 |    2 +
>  arch/arm/boot/dts/imx6sl.dtsi                  |    2 +
>  arch/arm/include/asm/hardware/cache-l2x0.h     |   63 +-
>  arch/arm/include/asm/outercache.h              |   67 +-
>  arch/arm/mach-cns3xxx/core.c                   |    8 +-
>  arch/arm/mach-highbank/highbank.c              |   12 +-
>  arch/arm/mach-imx/system.c                     |    8 +-
>  arch/arm/mach-omap2/omap-mpuss-lowpower.c      |    2 +-
>  arch/arm/mach-omap2/omap4-common.c             |   58 +-
>  arch/arm/mach-prima2/l2x0.c                    |    5 +-
>  arch/arm/mach-prima2/pm.c                      |    1 -
>  arch/arm/mach-realview/realview_pbx.c          |    4 +-
>  arch/arm/mach-spear/platsmp.c                  |   19 +-
>  arch/arm/mach-spear/spear13xx.c                |    6 +-
>  arch/arm/mach-sti/board-dt.c                   |    8 +-
>  arch/arm/mach-tegra/sleep.h                    |    8 +-
>  arch/arm/mach-ux500/cache-l2x0.c               |   23 +-
>  arch/arm/mach-vexpress/ct-ca9x4.c              |    4 +-
>  arch/arm/mm/Kconfig                            |   51 +
>  arch/arm/mm/Makefile                           |    1 +
>  arch/arm/mm/cache-feroceon-l2.c                |    1 -
>  arch/arm/mm/cache-l2x0.c                       | 1354 +++++++++++++++++-------
>  arch/arm/mm/l2c-common.c                       |   20 +
>  arch/arm/plat-samsung/s5p-sleep.S              |    8 +-
>  26 files changed, 1221 insertions(+), 567 deletions(-)
> 


-- 
Gregory Clement, Free Electrons
Kernel, drivers, real-time and embedded Linux
development, consulting, training and support.
http://free-electrons.com