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

[Shaohua Li]

On Fri, 2011-10-28 at 07:34 +0800, Minchan Kim wrote:
> On Tue, Oct 25, 2011 at 10:59:40AM +0800, Shaohua Li wrote:
> > With current logic, if page reclaim finds a huge page, it will just reclaim
> > the head page and leave tail pages reclaimed later. Let's take an example,
> > lru list has page A and B, page A is huge page:
> > 1. page A is isolated
> > 2. page B is isolated
> > 3. shrink_page_list() adds page A to swap page cache. so page A is split.
> > page A+1, page A+2, ... are added to lru list.
> > 4. shrink_page_list() adds page B to swap page cache.
> > 5. page A and B is written out and reclaimed.
> > 6. page A+1, A+2 ... is isolated and reclaimed later.
> > So the reclaim order is A, B, ...(maybe other pages), A+1, A+2 ...
> 
> I don't see your code yet but have a question.
> You mitigate this problem by 4/5 which could add subpages into lru tail
> so subpages would reclaim next interation of reclaim.
> 
> What do we need 5/5?
> Do I miss something?
Both patches are required. without this patch, current page reclaim will
only reclaim the first page of a huge page, because the hugepage isn't
split yet. The hugepage is split when the first page is being written to
swap, which is too later and page reclaim might already isolated a lot
of pages.
 
> > We expected the whole huge page A is reclaimed in the meantime, so
> > the order is A, A+1, ... A+HPAGE_PMD_NR-1, B, ....
> > 
> > With this patch, we do huge page split just after the head page is isolated
> > for inactive lru list, so the tail pages will be reclaimed immediately.
> > 
> > In a test, a range of anonymous memory is written and will trigger swap.
> > Without the patch:
> > #cat /proc/vmstat|grep thp
> > thp_fault_alloc 451
> > thp_fault_fallback 0
> > thp_collapse_alloc 0
> > thp_collapse_alloc_failed 0
> > thp_split 238
> > 
> > With the patch:
> > #cat /proc/vmstat|grep thp
> > thp_fault_alloc 450
> > thp_fault_fallback 1
> > thp_collapse_alloc 0
> > thp_collapse_alloc_failed 0
> > thp_split 103
> > 
> > So the thp_split number is reduced a lot, though there is one extra
> > thp_fault_fallback.
> 
> Wow. The result seems to be good.
> Is it result of effect only 5/5? or both 4/5 and 5/5?
both are required.

--
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/