Re: [PATCH RFC] bootmem: micro optimize freeing pages in bulks

[Johannes Weiner <hannes@cmpxchg.org>]

On Tue, Dec 13, 2011 at 11:06:01AM +0100, Johannes Weiner wrote:
> On Thu, Dec 01, 2011 at 11:10:55PM +0100, Uwe Kleine-Konig wrote:
> > The first entry of bdata->node_bootmem_map holds the data for
> > bdata->node_min_pfn up to bdata->node_min_pfn + BITS_PER_LONG - 1. So
> > the test for freeing all pages of a single map entry can be slightly
> > relaxed.
> 
> Agreed.  The optimization is tiny - we may lose one bulk order-5/6
> free per node and do it in 32/64 order-0 frees instead (we touch each
> page anyway one way or another), but the code makes more sense with
> your change.
> 
> [ Btw, what's worse is start being unaligned, because we won't do a
>   single batch free then.  The single-page loop should probably just
>   move to the next BITS_PER_WORD boundary and then retry the aligned
>   batch frees.  Oh, well... ]
> 
> > Moreover use DIV_ROUND_UP in another place instead of open coding it.
> 
> Agreed.
> 
> > Cc: Johannes Weiner <hannes@saeurebad.de>
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Signed-off-by: Uwe Kleine-Konig <u.kleine-koenig@pengutronix.de>
> > ---
> > Hello,
> > 
> > I'm not sure the current code is correct (and my patch doesn't fix it):
> > 
> > If
> > 
> > 	aligned && vec == ~0UL
> > 
> > evalutates to true, but
> > 
> > 	start + BITS_PER_LONG <= end
> > 
> > does not (or "< end" resp.) the else branch still frees all BITS_PER_LONG
> > pages. Is this intended? If yes, the last check can better be omitted
> > resulting in the pages being freed in a bulk.
> > If not, the loop in the else branch should only do something like:
> > 
> > 	while (vec && off < min(BITS_PER_LONG, end - start)) {
> > 		...
> 
> I would think this is fine because node_bootmem_map, which is where
> vec points to, is sized in multiples of pages, and zeroed word-wise.
> So even if end is not aligned, we can rely on !vec.

And putting two and two together (watch out, I'm on fire!) for the
other branch, this means that the whole

	start + BITS_PER_LONG <= end

check is actually unnecessary.  Your patch is still correct, of
course, but we could as well just remove the whole thing and rely on
vec not having bits set for the address range past end.

> > @@ -197,7 +197,7 @@ static unsigned long __init free_all_bootmem_core(bootmem_data_t *bdata)
> >  		idx = start - bdata->node_min_pfn;
> >  		vec = ~map[idx / BITS_PER_LONG];
> >  
> > -		if (aligned && vec == ~0UL && start + BITS_PER_LONG < end) {
> > +		if (aligned && vec == ~0UL && start + BITS_PER_LONG <= end) {
> >  			int order = ilog2(BITS_PER_LONG);
> >  
> >  			__free_pages_bootmem(pfn_to_page(start), order);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>