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

From: Minchan Kim
Date: Thu Aug 15 2013 - 00:18:19 EST


Hello,

On Thu, Aug 15, 2013 at 11:46:07AM +0800, Xishi Qiu wrote:
> On 2013/8/15 10:44, Minchan Kim wrote:
>
> > 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
>
> Hi Minchan,
> I mean if the private is set to a large number, it will skip 2^private
> pages, not (pageblock_nr_pages -1). I find somewhere will use page->private,
> such as fs. Here is the comment about parivate.
> /* Mapping-private opaque data:
> * usually used for buffer_heads
> * if PagePrivate set; used for
> * swp_entry_t if PageSwapCache;
> * indicates order in the buddy
> * system if PG_buddy is set.
> */

Please read full thread in detail.

Mel suggested following as

if (PageBuddy(page)) {
int nr_pages = (1 << page_order(page)) - 1;
if (PageBuddy(page)) {
nr_pages = min(nr_pages, MAX_ORDER_NR_PAGES - 1);
low_pfn += nr_pages;
continue;
}
}

min(nr_pages, xxx) removes your concern but I think Mel's version
isn't right. It should be aligned with pageblock boundary so I
suggested following.

if (PageBuddy(page)) {
#ifdef CONFIG_MEMORY_ISOLATION
unsigned long order = page_order(page);
if (PageBuddy(page)) {
low_pfn += (1 << order) - 1;
low_pfn = min(low_pfn, end_pfn);
}
#endif
continue;
}

so worst case is (pageblock_nr_pages - 1).
but we don't need to add CONFIG_MEMORY_ISOLATION so my suggestion
is following as.

if (PageBuddy(page)) {
unsigned long order = page_order(page);
if (PageBuddy(page)) {
low_pfn += (1 << order) - 1;
low_pfn = min(low_pfn, end_pfn);
}
continue;
}


> Thanks,
> Xishi Qiu
>
> > 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/