Re: [PATCH 16/22] Do not setup zonelist cache when there is only one node

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

> On Wed, 22 Apr 2009, Lee Schermerhorn wrote:
> 
> > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > > index 7f45de1..e59bb80 100644
> > > > --- a/mm/page_alloc.c
> > > > +++ b/mm/page_alloc.c
> > > > @@ -1467,8 +1467,11 @@ this_zone_full:
> > > >  		if (NUMA_BUILD)
> > > >  			zlc_mark_zone_full(zonelist, z);
> > > 
> > > If zonelist caching is never used for UMA machines, why should they ever 
> > > call zlc_mark_zone_full()?  It will always dereference 
> > > zonelist->zlcache_ptr and immediately return without doing anything.
> > > 
> > > Wouldn't it better to just add
> > > 
> > > 	if (num_online_nodes() == 1)
> > > 		continue;
> > > 
> > > right before this call to zlc_mark_zone_full()?  This should compile out 
> > > the remainder of the loop for !CONFIG_NUMA kernels anyway.
> > 
> > Shouldn't it already do that?  NUMA_BUILD is defined as 0 when
> > !CONFIG_NUMA to avoid #ifdef's in the code while still allowing compiler
> > error checking in the dead code.
> > 
> 
> Yeah, but adding the check on num_online_nodes() also prevents needlessly 
> calling zlc_mark_zone_full() on CONFIG_NUMA kernels when running on an UMA 
> machine.

I don't like this idea...

In UMA system, zlc_mark_zone_full() isn't so expensive. but In large system
one branch increasing is often costly.