Re: [PATCH v2] mm/gup: fix try_grab_compound_head() race with split_huge_page()

[Jason Gunthorpe]

On Tue, Jun 15, 2021 at 02:09:38PM +0200, Jann Horn wrote:
> On Tue, Jun 15, 2021 at 8:37 AM John Hubbard <jhubbard@xxxxxxxxxx> wrote:
> > On 6/14/21 6:20 PM, Jann Horn wrote:
> > > try_grab_compound_head() is used to grab a reference to a page from
> > > get_user_pages_fast(), which is only protected against concurrent
> > > freeing of page tables (via local_irq_save()), but not against
> > > concurrent TLB flushes, freeing of data pages, or splitting of compound
> > > pages.
> [...]
> > Reviewed-by: John Hubbard <jhubbard@xxxxxxxxxx>
> 
> Thanks!
> 
> [...]
> > > @@ -55,8 +72,23 @@ static inline struct page *try_get_compound_head(struct page *page, int refs)
> > >       if (WARN_ON_ONCE(page_ref_count(head) < 0))
> > >               return NULL;
> > >       if (unlikely(!page_cache_add_speculative(head, refs)))
> > >               return NULL;
> > > +
> > > +     /*
> > > +      * At this point we have a stable reference to the head page; but it
> > > +      * could be that between the compound_head() lookup and the refcount
> > > +      * increment, the compound page was split, in which case we'd end up
> > > +      * holding a reference on a page that has nothing to do with the page
> > > +      * we were given anymore.
> > > +      * So now that the head page is stable, recheck that the pages still
> > > +      * belong together.
> > > +      */
> > > +     if (unlikely(compound_head(page) != head)) {
> >
> > I was just wondering about what all could happen here. Such as: page gets split,
> > reallocated into a different-sized compound page, one that still has page pointing
> > to head. I think that's OK, because we don't look at or change other huge page
> > fields.
> >
> > But I thought I'd mention the idea in case anyone else has any clever ideas about
> > how this simple check might be insufficient here. It seems fine to me, but I
> > routinely lack enough imagination about concurrent operations. :)
> 
> Hmmm... I think the scariest aspect here is probably the interaction
> with concurrent allocation of a compound page on architectures with
> store-store reordering (like ARM). *If* the page allocator handled
> compound pages with lockless, non-atomic percpu freelists, I think it
> might be possible that the zeroing of tail_page->compound_head in
> put_page() could be reordered after the page has been freed,
> reallocated and set to refcount 1 again?

Oh wow, yes, this all looks sketchy! Doing a RCU access to page->head
is a really challenging thing :\

On the simplified store side:

  page->head = my_compound
  *ptep = page

There must be some kind of release barrier between those two
operations or this is all broken.. That definately deserves a comment.

Ideally we'd use smp_store_release to install the *pte :\

Assuming we cover the release barrier, I would think the algorithm
should be broadly:

 struct page *target_page = READ_ONCE(pte)
 struct page *target_folio = READ_ONCE(target_page->head)

 page_cache_add_speculative(target_folio, refs)

 if (target_folio != READ_ONCE(target_page->head) ||
     target_page != READ_ONCE(pte))
    goto abort

Which is what this patch does but I would like to see the
READ_ONCE's.

And there possibly should be two try_grab_compound_head()'s since we
don't need this overhead on the fully locked path, especially the
double atomic on page_ref_add()

> I think the lockless page cache code also has to deal with somewhat
> similar ordering concerns when it uses page_cache_get_speculative(),
> e.g. in mapping_get_entry() - first it looks up a page pointer with
> xas_load(), and any access to the page later on would be a _dependent
> load_, but if the page then gets freed, reallocated, and inserted into
> the page cache again before the refcount increment and the re-check
> using xas_reload(), then there would be no data dependency from
> xas_reload() to the following use of the page...

xas_store() should have the smp_store_release() inside it at least..

Even so it doesn't seem to do page->head, so this is not quite the
same thing

Jason