Re: + mm-memblock-reduce-overhead-in-binary-search.patch added to -mm tree

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 8 Oct 2012 12:42:34 -0700
Andrew Morton <akpm@linux-foundation.org> wrote:

> On Mon, 10 Sep 2012 13:55:15 +0200
> Michal Hocko <mhocko@suse.cz> wrote:
> 
> > > >OK. Thanks for the clarification. The main question remains, though. Is
> > > >this worth for memblock_is_memory?
> > > 
> > > There are many call sites need to call pfn_valid, how can you guarantee all
> > > the addrs are between memblock_start_of_DRAM() and memblock_end_of_DRAM(), 
> > > if not can this reduce possible overhead ? 
> > 
> > That was my question. I hoped for an answer in the patch description. I
> > am really not familiar with unicore32 which is the only user now.
> > 
> > > I add unlikely which means that this will not happen frequently. :-)
> > 
> > unlikely doesn't help much in this case. You would be doing the test for
> > every pfn_valid invocation anyway. So the main question is. Do you want
> > to optimize for something that doesn't happen often when it adds a cost
> > (not a big one but still) for the more probable cases?
> > I would say yes if we clearly see that the exceptional case really pays
> > off. Nothing in the changelog convinces me about that.
> 
> I don't believe Michal's questions have been resolved yet, so I'll keep
> this patch on hold for now.

ETIMEDOUT.  I'll drop the patch.  Please resend if you think it's still
needed and if these questions can be addressed.