Re: [Qestion] Hit a WARN_ON_ONCE in try_to_unmap_one when runing syzkaller

[Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>]

Hi,

On Wed, Mar 13, 2019 at 12:03:20AM +0800, zhong jiang wrote:
...
> 
> Minchan has changed the conditon check from  BUG_ON  to WARN_ON_ONCE in try_to_unmap_one.
> However,  It is still an abnormal condition when PageSwapBacked is not equal to PageSwapCache.
> 
> But Is there any case it will meet the conditon in the mainline.
> 
> It is assumed that PageSwapBacked(page) is true in the anonymous page,   This is to say,  PageSwapcache
> is false. however,  That is impossible because we will update the pte for hwpoison entry.
> 
> Because page is locked ,  Its page flags should not be changed except for PageSwapBacked

try_to_unmap_one() from hwpoison_user_mappings() could reach the
WARN_ON_ONCE() only if TTU_IGNORE_HWPOISON is set, because PageHWPoison()
is set at the beginning of memory_failure().

Clearing TTU_IGNORE_HWPOISON might happen on the following two paths:

  static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
                                    int flags, struct page **hpagep)
  {
      ...
  
      if (PageSwapCache(p)) {
              pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
                      pfn);
              ttu |= TTU_IGNORE_HWPOISON;
      }
      ...

      mapping = page_mapping(hpage);                                                                           
      if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&                                           
          mapping_cap_writeback_dirty(mapping)) {                                                              
              if (page_mkclean(hpage)) {                                                                       
                      SetPageDirty(hpage);                                                                     
              } else {                                                                                         
                      kill = 0;                                                                                
                      ttu |= TTU_IGNORE_HWPOISON;                                                              
                      pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
                              pfn);                                                                            
              }                                                                                                
      }                                                                                                        
      ...

      unmap_success = try_to_unmap(hpage, ttu);
      ...

So either of the above "ttu |= TTU_IGNORE_HWPOISON" should be executed.
I'm not sure which one, but both paths show printk messages, so if you
could have kernel message log, that might help ...

Thanks,
Naoya Horiguchi