[PATCH v2 05/23] ARM: Kirkwood: Seperate board-dt from common and pcie code.

[jgunthorpe@obsidianresearch.com (Jason Gunthorpe)]

On Sat, Feb 15, 2014 at 03:05:05PM +0100, Andrew Lunn wrote:
> On Sat, Feb 15, 2014 at 02:18:08PM +0100, Arnd Bergmann wrote:
> > On Saturday 15 February 2014 11:20:03 Andrew Lunn wrote:
> > > +static void __init kirkwood_l2_init(void)
> > > +{
> > > +#ifdef CONFIG_CACHE_FEROCEON_L2
> > > +#ifdef CONFIG_CACHE_FEROCEON_L2_WRITETHROUGH
> > > +       writel(readl(L2_CONFIG_REG) | L2_WRITETHROUGH, L2_CONFIG_REG);
> > > +       feroceon_l2_init(1);
> > > +#else
> > > +       writel(readl(L2_CONFIG_REG) & ~L2_WRITETHROUGH, L2_CONFIG_REG);
> > > +       feroceon_l2_init(0);
> > > +#endif
> > > +#endif
> > > +}
> > 
> > I assume this is correct, but I don't understand it. Why is there a
> > configuration option for this? Do both write-through and write-back
> > work on all machines, or could there be a case where some machine
> > actually requires a particular mode? If not, isn't write-back
> > normally "better", so you won't actually ever want to set write-through
> > mode?
> 
> Hi Arnd
> 
> The honest answer is, i've no idea. I'm just shuffling code around in
> this patch, and not applying my brain as to what this code does....
> 
> Maybe JasonG has a better idea of this history of this?

No idea of the history, but for DMA heavy work loads write-through is
better since you spend less cpu cycles doing cache flushing, while for
CPU centric work loads write-back is better since you spend less time
waiting for memory.

Since these SOCs are popular for storage and networking apps I'm not
surprised to see this option.

But a static config option is not really in-line with current thinking
on these sorts of things. A DT option would be better (IMHO), but even
that is probably not going to be universally loved.

Can we worry about this after Andrew's shuffling is done?

Regards,
Jason