Re: split_huge_page_to_list() races with page_mapcount() on migration entry in smaps code? [was: Re: [syzbot] kernel BUG in __page_mapcount]

[Matthew Wilcox]

On Mon, Jun 07, 2021 at 09:55:09PM +0200, Jann Horn wrote:
> On Mon, Jun 7, 2021 at 8:03 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > On Mon, Jun 07, 2021 at 07:27:23PM +0200, Jann Horn wrote:
> > > === Short summary ===
> > > I believe the issue here is a race between /proc/*/smaps and
> > > split_huge_page_to_list():
> > >
> > > The codepath for /proc/*/smaps walks the pagetables and (e.g. in
> > > smaps_account()) calls page_mapcount() not just on pages from normal
> > > PTEs but also on migration entries (since commit b1d4d9e0cbd0a
> > > "proc/smaps: carefully handle migration entries", from Linux v3.5).
> > > page_mapcount() expects compound pages to be stable.
> > >
> > > The split_huge_page_to_list() path first protects the compound page by
> > > locking it and replacing all its PTEs with migration entries (since
> > > the THP rewrite in v4.5, I think?), then does the actual splitting
> > > using __split_huge_page().
> > >
> > > So there's a mismatch of expectations here:
> > > The smaps code expects that migration entries point to stable compound
> > > pages, while the THP code expects that it's okay to split a compound
> > > page while it has migration entries.
> >
> > Will it be a colossal performance penalty if we always get the page
> > refcount after looking it up?  That will cause split_huge_page() to
> > fail to split the page if it hits this race.
> 
> Hmm - but with that approach I'm not sure you could even easily take a
> refcount on a page whose refcount may be frozen and which may be in
> the middle of being shattered? get_page_unless_zero() is wrong because
> you can't take references on tail pages, right? (Or can you?) And
> try_get_page() is wrong because it bugs out if the refcount is zero -
> and even if it didn't do that, you might end up holding a reference on
> the head page while the page you're actually interested in is a tail
> page?
> 
> I guess if it was really necessary, it'd be possible to do some kind
> of retry thing that grabs a reference on the compound head, then
> checks that the tail page is still associated with the compound head,
> and if not, drops the compound head and tries again?

Right; that's how get_user_page_fast() works -- see
try_get_compound_head().  If it can't get the reference, it just fails.
I suspect for smaps, we can just choose to not count the page.  It'll be
an inaccuracy in the stats, but I don't think that's a big deal.