Re: [f2fs-dev] [syzbot] BUG: unable to handle kernel NULL pointer dereference in folio_mark_dirty

[Jaegeuk Kim <jaegeuk@kernel.org>]

On 12/07, Matthew Wilcox wrote:
> On Tue, Dec 07, 2021 at 01:39:06PM -0800, Jaegeuk Kim wrote:
> > On 12/07, Matthew Wilcox wrote:
> > > > > Call Trace:
> > > > >  <TASK>
> > > > >  folio_mark_dirty+0x136/0x270 mm/page-writeback.c:2639
> > > 
> > >         if (likely(mapping)) {
> > > ...
> > >                 if (folio_test_reclaim(folio))
> > >                         folio_clear_reclaim(folio);
> > >                 return mapping->a_ops->set_page_dirty(&folio->page);
> > > 
> > > how do we get to a NULL ->set_page_dirty for a metadata page's
> > > mapping->a_ops?  This is definitely an f2fs expert question.
> > 
> > I can't find anything in f2fs, since that page was got by f2fs_grab_meta_page
> > along with grab_cache_page() that we never unlocked it.
> > 
> >   40 struct page *f2fs_grab_meta_page(struct f2fs_sb_info *sbi, pgoff_t index)
> >   41 {
> >   42         struct address_space *mapping = META_MAPPING(sbi);
> >   43         struct page *page;
> >   44 repeat:
> >   45         page = f2fs_grab_cache_page(mapping, index, false);
> > 
> >                     -> grab_cache_page(mapping, index);
> > 
> >   46         if (!page) {
> >   47                 cond_resched();
> >   48                 goto repeat;
> >   49         }
> >   50         f2fs_wait_on_page_writeback(page, META, true, true);
> >   51         if (!PageUptodate(page))
> >   52                 SetPageUptodate(page);
> >   53         return page;
> >   54 }
> > 
> > 
> > Suspecting something in folio wrt folio_mapping()?
> > 
> >  81 bool set_page_dirty(struct page *page)
> >  82 {
> >  83         return folio_mark_dirty(page_folio(page));
> >  84 }
> 
> ... huh?  How could folio_mapping() be getting this wrong?

Dunno.

> page_folio() does the same thing as compound_head() -- as far as I know
> you don't use compound pages for f2fs metadata, so this basically just
> casts the page to a struct folio.
> 
> folio_mapping() is just like the old page_mapping() (see commit
> 2f52578f9c64).  Unless you've done something like set the swapcache
> bit on your metadata page, it's just going to return folio->mapping
> (ie the same as page->mapping).

Hmm, I've never seen this call stack before, so simply started to suspect
folio.