[PATCH] ARM: implement optimized percpu variable access

[Will Deacon]

On Fri, Nov 23, 2012 at 05:06:07PM +0000, Rob Herring wrote:
> On 11/22/2012 05:34 AM, Will Deacon wrote:
> > On Sun, Nov 11, 2012 at 03:20:40AM +0000, Rob Herring wrote:
> >> From: Rob Herring <rob.herring at calxeda.com>
> >>
> >> Use the previously unused TPIDRPRW register to store percpu offsets.
> >> TPIDRPRW is only accessible in PL1, so it can only be used in the kernel.
> >>
> >> This saves 2 loads for each percpu variable access which should yield
> >> improved performance, but the improvement has not been quantified.
> >>
> >> Signed-off-by: Rob Herring <rob.herring at calxeda.com>
> >> ---
> >>  arch/arm/include/asm/Kbuild   |    1 -
> >>  arch/arm/include/asm/percpu.h |   44 +++++++++++++++++++++++++++++++++++++++++
> >>  arch/arm/kernel/smp.c         |    3 +++
> >>  3 files changed, 47 insertions(+), 1 deletion(-)
> >>  create mode 100644 arch/arm/include/asm/percpu.h
> > 
> > Russell pointed out to me that this patch will break on v6 CPUs if they don't
> > have the thread ID registers and we're running with SMP_ON_UP=y. Looking at
> > the TRMs, the only case we care about is 1136 < r1p0, but it does indeed break
> > there (I have a board on my desk).
> 
> Are there any non ARM Ltd. cores without v6K we need to worry about? I
> wouldn't think there are many 1136 < r1p0 out there (your desk being an
> obvious exception).

To be honest, I'm not sure. It would be good if Marvell and Qualcomm could
chime in as I wouldn't be surprised if they had some parts that fit this
category.

As for the 1136, we have a few spare if you want one!

> > There are a few ways to fix this:
> > 
> > (1) Use the SMP alternates to patch the code when running on UP systems. I
> >     tried this and the code is absolutely diabolical (see below).
> > 
> > (2) Rely on the registers being RAZ/WI rather than undef (which seems to be
> >     the case on my board) and add on the pcpu delta manually. This is also
> >     really horrible.
> > 
> > (3) Just make the thing depend on __LINUX_ARM_ARCH__ >= 7. Yes, we lose on
> >     11MPCore, but we win on A8 and the code is much, much simpler.
> 
> I would lean towards this option. It really only has to depend on v6K
> and !v6. We can refine the multi-platform selection to allow v7 and v6K
> only builds in addition to v7 and v6. I think we are going to have
> enough optimizations with v7 (gcc optimizations, thumb2, unaligned
> accesses, etc.) that most people will do v7 only builds.

Sounds alright to me, and I'm happy to test on the boards that I have.

Will