[PATCH 5/7] [RFC] force 32-byte aligned kmallocs

[Benjamin Herrenschmidt]

On Thu, 2011-05-19 at 19:47 -0500, Eric Van Hensbergen wrote:
> On Thu, May 19, 2011 at 7:36 PM, Benjamin Herrenschmidt
> <benh at kernel.crashing.org> wrote:
> > On Wed, 2011-05-18 at 16:24 -0500, Eric Van Hensbergen wrote:
> >>
> >> -#ifdef CONFIG_NOT_COHERENT_CACHE
> >> +#if defined(CONFIG_NOT_COHERENT_CACHE) || defined(CONFIG_BGP)
> >>  #define ARCH_DMA_MINALIGN    L1_CACHE_BYTES
> >>  #endif
> >
> > Is DMA cache coherent on BG/P ? That's odd for a 4xx base :-)
> >
> 
> My understanding of things (which could be totally wrong) is that the
> DMA we care about on BG/P (namely the Torus and Collective networks)
> is coherent at the L2.  Of course the change in question is talking
> about L1_CACHE_BYTES, so my reading of this is that its a sleazy way
> of getting aligned mallocs that make interactions with the tightly
> coupled networks easier/more-efficient.  I'm open to alternative
> suggestions.

But if it's not coherent with L1, then you sould have
CONFIG_NOT_COHERENT_CACHE set and not need that patch... or am I missing
something ?

One thing we should do some day as well is make that whole non-coherent
be runtime selected, on the list of things to fix 440+47x in the same
kernel. Pfiew....

Cheers,
Ben.