Re: [patch 5/5]thp: split huge page if head page is isolated

[Minchan Kim]

On Thu, Nov 10, 2011 at 10:46:07AM +0800, Shaohua Li wrote:
> On Thu, 2011-11-10 at 10:23 +0800, Minchan Kim wrote:
> > So long contents.
> > Let's remove it.
> > 
> > On Thu, Nov 10, 2011 at 10:07:10AM +0800, Shaohua Li wrote:
> > 
> > <snip>
> > 
> > > > > Coudn't we make both sides good?
> > > > > 
> > > > > Here is my quick patch.
> > > > > How about this?
> > > > > It doesn't split THPs in page_list but still reclaims non-THPs so
> > > > > I think it doesn't changed old behavior a lot.
> > > > I like this idea, will do some test soon.
> > > hmm, this doesn't work as expected. The putback_lru_page() messes lru.
> > > This isn't a problem if the page will be written since
> > > rotate_reclaimable_page() will fix the order. I got worse data than my
> > > v2 patch, eg, more thp_fallbacks, mess lru order, more pages are
> > > scanned. We could add something like putback_lru_page_tail, but I'm not
> > 
> > Hmm, It's not LRU mess problem. but it's just guessing and you might be right
> > because you have a workload and can test it.
> > 
> > My guessing is that cull_mlocked reset synchronus page reclaim.
> > Could you test this patch, again?
> no, I traced it, and lru mess. putback_lru_page() adds the page to lru
> head instead of tail.

I knew LRU mess happens but I mean I am not sure the culprit is it.

> 
> > And, if the problem cause by LRU mess, I think it is valuable with adding putback_lru_page_tail
> > because thp added lru_add_page_tail, too.
> I want to put all remaining pages back to lru tail if a huge page is
> split, because enough pages are reclaimed. So this needs adding
> something like putback_lru_pages_tail(), not complicated, but a lot of
> code. And if there are parallel reclaimer, we still have lru mess. My
> test already shows it. Still worthy?

If parallel reclaim happens, it can spoil everything. It's really really bad.
I perfer adding putback_lru_page_tail with isoloation trick because probably
we can use it later about pages which was not able to reclaim by some causes temporally.

> 
> Thanks,
> Shaohua
> 

-- 
Kind regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/