Re: [PATCH] mm: skip the page buddy block instead of one page

[Minchan Kim]

Hi Xishi,

On Thu, Aug 15, 2013 at 10:32:50AM +0800, Xishi Qiu wrote:
> On 2013/8/15 2:00, Mel Gorman wrote:
> 
> >>> Even if the page is still page buddy, there is no guarantee that it's
> >>> the same page order as the first read. It could have be currently
> >>> merging with adjacent buddies for example. There is also a really
> >>> small race that a page was freed, allocated with some number stuffed
> >>> into page->private and freed again before the second PageBuddy check.
> >>> It's a bit of a hand grenade. How much of a performance benefit is there
> >>
> >> 1. Just worst case is skipping pageblock_nr_pages
> > 
> > No, the worst case is that page_order returns a number that is
> > completely garbage and low_pfn goes off the end of the zone
> > 
> >> 2. Race is really small
> >> 3. Higher order page allocation customer always have graceful fallback.
> >>
> 
> Hi Minchan, 
> I think in this case, we may get the wrong value from page_order(page).
> 
> 1. page is in page buddy
> 
> > if (PageBuddy(page)) {
> 
> 2. someone allocated the page, and set page->private to another value
> 
> > 	int nr_pages = (1 << page_order(page)) - 1;
> 
> 3. someone freed the page
> 
> > 	if (PageBuddy(page)) {
> 
> 4. we will skip wrong pages

So, what's the result by that?
As I said, it's just skipping (pageblock_nr_pages -1) at worst case
and the case you mentioned is right academically and I and Mel
already pointed out that. But how often could that happen in real
practice? I believe such is REALLY REALLY rare.
So, as Mel said, if you have some workloads to see the benefit
from this patch, I think we could accept the patch.
Could you try and respin with the number?
I guess big contigous memory range or memory-hotplug which are
full of free pages in embedded CPU which is rather slower than server
or desktop side could have benefit.

Thanks.

> 
> > 		nr_pages = min(nr_pages, MAX_ORDER_NR_PAGES - 1);
> > 		low_pfn += nr_pages;
> > 		continue;
> > 	}
> > }
> > 
> > It's still race-prone meaning that it really should be backed by some
> > performance data justifying it.
> > 
> 
> 
> 

-- 
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/