[PATCH] arm: allow all architectures to change max zone order

[Russell King - ARM Linux]

On Mon, Jul 01, 2013 at 04:33:49PM +0200, Lucas Stach wrote:
> Am Freitag, den 31.05.2013, 19:02 +0200 schrieb Lucas Stach:
> > Many multimedia applications on architectures without an
> > IOMMU need large physically contiguous buffers and so need
> > to adjust the max zone order of the page allocator.
> > 
> > There is zero reason to not allow this adjustment on arches
> > other than SHMOBILE.
> > 
> > Signed-off-by: Lucas Stach <l.stach at pengutronix.de>
> > ---
> >  arch/arm/Kconfig |    2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> > index 49d993c..6ceace8 100644
> > --- a/arch/arm/Kconfig
> > +++ b/arch/arm/Kconfig
> > @@ -1710,7 +1710,7 @@ config HW_PERF_EVENTS
> >  source "mm/Kconfig"
> >  
> >  config FORCE_MAX_ZONEORDER
> > -	int "Maximum zone order" if ARCH_SHMOBILE
> > +	int "Maximum zone order"
> >  	range 11 64 if ARCH_SHMOBILE
> >  	default "12" if SOC_AM33XX
> >  	default "9" if SA1111
> 
> Ping. This is really straight forward and low risk.

According to my greps, this is the only copy of this mail that appeared
on the mailing lists in 2013.  I've really not been happy with the
blatent assertion in this commit message, so I've been ignoring the
patch in the patch system because of that... until now.

First, let's look at the default, which is 11.  That corresponds with
2^10 pages, which gives a maximum allocation of 4MB.  While that's
rather limiting when you want to allocate a 1080p framebuffer (which
works out at slightly over 8MB), it seems that it's entirely possible
to allocate such framebuffers without resorting to fiddling with this
when CMA is being used.

Indeed, on iMX6 with imx-drm, I have a 1080p frame buffer provided by
the current CMA memory backed implementation there just fine with the
default value of this.  So, large physically contiguous buffers can
definitely be allocated today without problem - with CMA.

Not using CMA for large memory buffers would be a mistake.  Memory
suffers from fragmentation, which makes large memory allocations
suffer - I see even 16K allocations fail with 3.x kernels.  So what
hope has a 4MB (or larger) allocation got?

CMA avoids that by setting aside a special region which is only used
for data which can be moved into other memory areas when required,
thereby reducing the effects of memory fragmentation in this area.

So, far from "zero reason", there's a very real reason: memory
fragmentation leading to OOM situations.  Rather than fiddling with
this seemingly easy-to-fiddle-with constant, please use an appropriately
adjusted CMA for your multimedia applications instead of alloc_pages().

Thanks.

-- 
FTTC broadband for 0.8mile line: now at 9.7Mbps down 460kbps up... slowly
improving, and getting towards what was expected from it.