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

[Ding Hui]

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@xxxxxxxxxxxxxx>
> > > ---
> > >    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.
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

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;
    }

    lock_page(page);
	


--
Thanks,
- Ding Hui