Re: [tip:x86/mm] x86/pat: Avoid contention on cpa_lock if possible

[Ido Yariv <ido@wizery.com>]

Hi,

On Wed, Jun 06, 2012 at 06:18:12PM -0400, Shai Fultheim (Shai@ScaleMP.com) wrote:
> Peter Zijlstra [mailto:a.p.zijlstra@chello.nl] wrote
> > On Wed, 2012-06-06 at 09:18 -0700, tip-bot for Shai Fultheim wrote:
> > 
> > > [ I absolutely hate these locking patterns ... yet I have no better idea.  Maybe the gents on Cc: ... ]
> > > Signed-off-by: Ingo Molnar <mingo@kernel.org>
> > 
> > Oh yuck, this is vile..
> > 
> > static struct static_key scale_mp_trainwreck = STATIC_KEY_INIT_FALSE;
> > 
> > static DEFINE_SPINLOCK(_cpa_lock);
> > 
> > static inline void cpa_lock(void)
> > {
> > 	if (static_key_false(&scale_mp_trainwreck))
> > 		return;
> > 
> > 	spin_lock(&_cpa_lock);
> > }
> > 
> > static inline void cpa_unlock(void)
> > {
> > 	if (static_key_false(&scale_mp_trainwreck))
> > 		return;
> > 
> > 	spin_lock(&_cpa_lock);
> > }
> > 
> > And then use cpa_{,un}lock(), and the scale-mp guys can
> > static_key_slow_inc(&scale_mp_trainwreck).
> > 
> > [ and yes I hate those jump_label names ... but I'm not wanting
> >   to go through another round of bike-shed painting. ]
> 
> Looks pretty straight forward to do.  
> We will try this route, as I'm concerned that synthetic CPUID bit will be kind of a global change for a pretty local consideration.
> 
> Comments?

The synthetic CPUID bit approach has the advantage of setting this bit
from the platform initialization code independently (in this case, in
vsmp_64.c) instead of calling platform specific functions from
pageattr.c (such as is_vsmp_box()) or exporting the static_key variable.

We'll give this a shot and send a revised patch after testing it.

Sounds good?

Thanks,
Ido.