Re: [RFC PATCH] mm/page_alloc: fix counting of free pages after take off from buddy

["HORIGUCHI NAOYA(堀口 直也)" <naoya.horiguchi@nec.com>]

On Thu, May 06, 2021 at 12:01:34PM +0800, Ding Hui wrote:
> On 2021/5/6 10:49, HORIGUCHI NAOYA(堀口 直也) wrote:
> > On Wed, Apr 28, 2021 at 04:54:59PM +0200, David Hildenbrand wrote:
> > > On 21.04.21 04:04, Ding Hui wrote:
> > > > Recently we found there is a lot MemFree left in /proc/meminfo after
> > > > do a lot of pages soft offline.
> > > > 
> > > > I think it's incorrect since NR_FREE_PAGES should not contain HWPoison pages.
> > > > After take_page_off_buddy, the page is no longer belong to buddy
> > > > allocator, and will not be used any more, but we maybe missed accounting
> > > > NR_FREE_PAGES in this situation.
> > > > 
> > > > Signed-off-by: Ding Hui <dinghui@sangfor.com.cn>
> > > > ---
> > > >    mm/page_alloc.c | 1 +
> > > >    1 file changed, 1 insertion(+)
> > > > 
> > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > > index cfc72873961d..8d65b62784d8 100644
> > > > --- a/mm/page_alloc.c
> > > > +++ b/mm/page_alloc.c
> > > > @@ -8947,6 +8947,7 @@ bool take_page_off_buddy(struct page *page)
> > > >    			del_page_from_free_list(page_head, zone, page_order);
> > > >    			break_down_buddy_pages(zone, page_head, page, 0,
> > > >    						page_order, migratetype);
> > > > +			__mod_zone_page_state(zone, NR_FREE_PAGES, -1);
> > > >    			ret = true;
> > > >    			break;
> > > >    		}
> > > > 
> > > 
> > > Should this use __mod_zone_freepage_state() instead?
> > 
> > Yes, __mod_zone_freepage_state() looks better to me.
> > 
> > And I think that maybe an additional __mod_zone_freepage_state() in
> > unpoison_memory() is necessary to cancel the decrement.  I thought of the
> > following, but it doesn't build because get_pfnblock_migratetype() is
> > available only in mm/page_alloc.c, so you might want to add a small exported
> > routine in mm/page_alloc.c and let it called from unpoison_memory().
> > 
> >    @@ -1899,8 +1899,12 @@ int unpoison_memory(unsigned long pfn)
> >            }
> >            if (!get_hwpoison_page(p, flags, 0)) {
> >    -               if (TestClearPageHWPoison(p))
> >    +               if (TestClearPageHWPoison(p)) {
> >    +                       int migratetype = get_pfnblock_migratetype(p, pfn);
> >    +
> >                            num_poisoned_pages_dec();
> >    +                       __mod_zone_freepage_state(page_zone(p), 1, migratetype);
> >    +               }
> >                    unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
> >                                     pfn, &unpoison_rs);
> >                    return 0;
> > 
> 
> I think there is another problem:
> In normal case, we keep the last refcount of the hwpoison page, so
> get_hwpoison_page should return 1. The NR_FREE_PAGES will be adjusted when
> call put_page.

I think that take_page_off_buddy() should not be called for this case
(the error page have remaining refcount).  So it seems that no need to
update NR_FREE_PAGES ?

> At race condition, we maybe leak the page because we does not put it back to
> buddy in unpoison_memory, however the HWPoison flag, num_poisoned_pages,
> NR_FREE_PAGES is adjusted correctly.
> 
> CPU0                        CPU1
> 
> soft_offline_page
>   soft_offline_free_page
>     page_handle_poison
>       take_page_off_buddy
>       SetPageHWPoison
>                             unpoison_memory
>                               if (!get_hwpoison_page(p))
>                                 TestClearPageHWPoison
>                                   num_poisoned_pages_dec
>                                 __mod_zone_freepage_state
>                                 return 0
>                                 /* miss put the page back to buddy */
>       page_ref_inc
>       num_poisoned_pages_inc

Thanks for checking this, unpoison_memory() is racy.  Recently we are suggesting
to introduce mf_mutex by [1].  Although this patch is not merged to mainline yet,
but it could be used to prevent the above race too.

[1] https://lore.kernel.org/linux-mm/20210427062953.2080293-2-nao.horiguchi@gmail.com/

> 
> How about do nothing and return -EBUSY (so the caller can retry) if unpoison
> a zero refcount page , or return 0 like 230ac719c500 ("mm/hwpoison: don't
> try to unpoison containment-failed pages") does ?
> 
>   @@ -1736,11 +1736,9 @@ int unpoison_memory(unsigned long pfn)
>     }
> 
>     if (!get_hwpoison_page(p, flags, 0)) {
>   -       if (TestClearPageHWPoison(p))
>   -           num_poisoned_pages_dec();
>   -       unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
>   +       unpoison_pr_info("Unpoison: Software-unpoisoned zero refcount page
> %#lx\n",
>   				 pfn, &unpoison_rs);
>   -       return 0;
>   +       return -EBUSY;

Currently unpoison_memory() does not work as reverse operation of take_page_off_buddy()
(it's simply broken), so implementing it at one time would be better.
I'll take time to fix unpoison_memory().

Thanks,
Naoya Horiguchi