Re: [PATCH 7/8] HWPOISON, hugetlb: fix unpoison for hugepage

[Naoya Horiguchi]

On Wed, Aug 25, 2010 at 10:54:32AM +0800, Wu Fengguang wrote:
> On Wed, Aug 25, 2010 at 07:55:26AM +0800, Naoya Horiguchi wrote:
> > Currently unpoisoning hugepages doesn't work because it's not enough
> > to just clear PG_HWPoison bits and we need to link the hugepage
> > to be unpoisoned back to the free hugepage list.
> > To do this, we get and put hwpoisoned hugepage whose refcount is 0.
> > 
> > Signed-off-by: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>
> > Signed-off-by: Jun'ichi Nomura <j-nomura@xxxxxxxxxxxxx>
> > ---
> >  mm/memory-failure.c |   16 +++++++++++++---
> >  1 files changed, 13 insertions(+), 3 deletions(-)
> > 
> > diff --git v2.6.36-rc2/mm/memory-failure.c v2.6.36-rc2/mm/memory-failure.c
> > index 60178d2..ab36690 100644
> > --- v2.6.36-rc2/mm/memory-failure.c
> > +++ v2.6.36-rc2/mm/memory-failure.c
> > @@ -1154,9 +1154,19 @@ int unpoison_memory(unsigned long pfn)
> >  	nr_pages = 1 << compound_order(page);
> >  
> >  	if (!get_page_unless_zero(page)) {
> > -		if (TestClearPageHWPoison(p))
> > +		/* The page to be unpoisoned was free one when hwpoisoned */
> > +		if (TestClearPageHWPoison(page))
> >  			atomic_long_sub(nr_pages, &mce_bad_pages);
> >  		pr_debug("MCE: Software-unpoisoned free page %#lx\n", pfn);
> > +		if (PageHuge(page)) {
> > +			/*
> > +			 * To unpoison free hugepage, we get and put it
> > +			 * to move it back to the free list.
> > +			 */
> > +			get_page(page);
> > +			clear_page_hwpoison_huge_page(page);
> > +			put_page(page);
> > +		}
> >  		return 0;
> >  	}
> 
> It's racy in free huge page detection.
> 
> alloc_huge_page() does not increase page refcount inside hugetlb_lock,
> the alloc_huge_page()=>alloc_buddy_huge_page() path even drops the
> lock temporarily! Then we never know reliably if a huge page is really
> free.

I agree.

> Here is a scratched fix. It is totally untested. Just want to notice
> you that with this patch, the huge page unpoisoning should go easier.

Great.
I adjusted this patch to real hugetlb code and passed libhugetlbfs test.

Thanks,
Naoya
--
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/