Re: [PATCH] rudimentary tracing for Classic RCU

[Andi Kleen <andi@firstfloor.org>]

On Thu, Oct 09, 2008 at 04:14:26PM +0900, KOSAKI Motohiro wrote:
> Hi Lai-san,
> 
> > >> In this case, why not define it as:
> > >> static char buf[20*NR_CPUS + 100];
> > > 
> > > Actually you should near never use NR_CPUS now but always num_possible_cpus()
> > > (or even num_online_cpus()) Using NR_CPUS can lead to extreme waste
> > > of memory on kernels which are compiled for 4096 CPUs for example.
> > > 
> > > And with num_possible_cpus() kmalloc is needed.
> > 
> > I thought the default value of NR_CPUS is 32.
> 
> it's pointless.
> Almost distribution use _very_ large NR_CPUS (likes 4096).

Right now it's more like 8-128, but at least there is some effort
to make a 4096 CPU distribution possible and Mike Travis has been
 tirelessly working on eradicating all the rogue NR_CPUs all over 
the tree.  So it's better to not add more.

> So, We should concern large NR_CPUS.

Agreed.

-Andi

-- 
ak@linux.intel.com