Re: [PATCH 05/28] mm, swap: sanitize swap cache lookup convention

[Barry Song <21cnbao@gmail.com>]

On Wed, May 21, 2025 at 7:10 AM Kairui Song <ryncsn@gmail.com> wrote:
>
> On Tue, May 20, 2025 at 12:41 PM Barry Song <21cnbao@gmail.com> wrote:
> >
> > On Tue, May 20, 2025 at 3:31 PM Kairui Song <ryncsn@gmail.com> wrote:
> > >
> > > On Mon, May 19, 2025 at 12:38 PM Barry Song <21cnbao@gmail.com> wrote:
> > > >
> > > > > From: Kairui Song <kasong@tencent.com>
> > > >
> > > > > diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
> > > > > index e5a0db7f3331..5b4f01aecf35 100644
> > > > > --- a/mm/userfaultfd.c
> > > > > +++ b/mm/userfaultfd.c
> > > > > @@ -1409,6 +1409,10 @@ static int move_pages_pte(struct mm_struct *mm, pmd_t *dst_pmd, pmd_t *src_pmd,
> > > > >                               goto retry;
> > > > >                       }
> > > > >               }
> > > > > +             if (!folio_swap_contains(src_folio, entry)) {
> > > > > +                     err = -EBUSY;
> > > > > +                     goto out;
> > > > > +             }
> > > >
> > > > It seems we don't need this. In move_swap_pte(), we have been checking pte pages
> > > > are stable:
> > > >
> > > >         if (!is_pte_pages_stable(dst_pte, src_pte, orig_dst_pte, orig_src_pte,
> > > >                                  dst_pmd, dst_pmdval)) {
> > > >                 double_pt_unlock(dst_ptl, src_ptl);
> > > >                 return -EAGAIN;
> > > >         }
> > >
> > > The tricky part is when swap_cache_get_folio returns the folio, both
> > > folio and ptes are unlocked. So is it possible that someone else
> > > swapped in the entries, then swapped them out again using the same
> > > entries?
> > >
> > > The folio will be different here but PTEs are still the same value to
> > > they will pass the is_pte_pages_stable check, we previously saw
> > > similar races with anon fault or shmem. I think more strict checking
> > > won't hurt here.
> >
> > This doesn't seem to be the same case as the one you fixed in
> > do_swap_page(). Here, we're hitting the swap cache, whereas in that
> > case, there was no one hitting the swap cache, and you used
> > swap_prepare() to set up the cache to fix the issue.
> >
> > By the way, if we're not hitting the swap cache, src_folio will be
> > NULL. Also, it seems that folio_swap_contains(src_folio, entry) does
> > not guard against that case either.
>
> Ah, that's true, it should be moved inside the if (folio) {...} block
> above. Thanks for catching this!
>
> > But I suspect we won't have a problem, since we're not swapping in —
> > we didn't read any stale data, right? Swap-in will only occur after we
> > move the PTEs.
>
> My concern is that a parallel swapin / swapout could result in the
> folio to be a completely irrelevant or invalid folio.
>
> It's not about the dst, but in the move src side, something like:
>
> CPU1                             CPU2
> move_pages_pte
>   folio = swap_cache_get_folio(...)
>     | Got folio A here
>   move_swap_pte
>                                  <swapin src_pte, using folio A>
>                                  <swapout src_pte, put folio A>
>                                    | Now folio A is no longer valid.
>                                    | It's very unlikely but here SWAP
>                                    | could reuse the same entry as above.


swap_cache_get_folio() does increment the folio's refcount, but it seems this
doesn't prevent do_swap_page() from freeing the swap entry after swapping
in src_pte with folio A, if it's a read fault.
for write fault, folio_ref_count(folio) == (1 + folio_nr_pages(folio))
will be false:

static inline bool should_try_to_free_swap(struct folio *folio,
                                           struct vm_area_struct *vma,
                                           unsigned int fault_flags)
{
       ...

        /*
         * If we want to map a page that's in the swapcache writable, we
         * have to detect via the refcount if we're really the exclusive
         * user. Try freeing the swapcache to get rid of the swapcache
         * reference only in case it's likely that we'll be the exlusive user.
         */
        return (fault_flags & FAULT_FLAG_WRITE) && !folio_test_ksm(folio) &&
                folio_ref_count(folio) == (1 + folio_nr_pages(folio));
}

and for swapout, __removing_mapping does check refcount as well:

static int __remove_mapping(struct address_space *mapping, struct folio *folio,
                            bool reclaimed, struct mem_cgroup *target_memcg)
{
        refcount = 1 + folio_nr_pages(folio);
        if (!folio_ref_freeze(folio, refcount))
                goto cannot_free;

}

However, since __remove_mapping() occurs after pageout(), it seems
this also doesn't prevent swapout from allocating a new swap entry to
fill src_pte.

It seems your concern is valid—unless I'm missing something.
Do you have a reproducer? If so, this will likely need a separate fix
patch rather than being hidden in this patchset.

>     double_pt_lock
>     is_pte_pages_stable
>       | Passed because of entry reuse.
>     folio_move_anon_rmap(...)
>       | Moved invalid folio A.
>
> And could it be possible that the swap_cache_get_folio returns NULL
> here, but later right before the double_pt_lock, a folio is added to
> swap cache? Maybe we better check the swap cache after clear and
> releasing dst lock, but before releasing src lock?

It seems you're suggesting that a parallel swap-in allocates and adds
a folio to the swap cache, but the PTE has not yet been updated from
a swap entry to a present mapping?

As long as do_swap_page() adds the folio to the swap cache
before updating the PTE to present, this scenario seems possible.

It seems we need to double-check:

diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
index bc473ad21202..976053bd2bf1 100644
--- a/mm/userfaultfd.c
+++ b/mm/userfaultfd.c
@@ -1102,8 +1102,14 @@ static int move_swap_pte(struct mm_struct *mm,
struct vm_area_struct *dst_vma,
        if (src_folio) {
                folio_move_anon_rmap(src_folio, dst_vma);
                src_folio->index = linear_page_index(dst_vma, dst_addr);
+       } else {
+               struct folio *folio =
filemap_get_folio(swap_address_space(entry),
+                                       swap_cache_index(entry));
+               if (!IS_ERR_OR_NULL(folio)) {
+                       double_pt_unlock(dst_ptl, src_ptl);
+                       return -EAGAIN;
+               }
        }
-
        orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
 #ifdef CONFIG_MEM_SOFT_DIRTY
        orig_src_pte = pte_swp_mksoft_dirty(orig_src_pte);

Let me run test case [1] to check whether this ever happens. I guess I need to
hack kernel a bit to always add folio to swapcache even for SYNC IO.

[1] https://lore.kernel.org/linux-mm/20250219112519.92853-1-21cnbao@gmail.com/

>
>
> >
> > >
> > > >
> > > > Also, -EBUSY is somehow incorrect error code.
> > >
> > > Yes, thanks, I'll use EAGAIN here just like move_swap_pte.
> > >
> > >
> > > >
> > > > >               err = move_swap_pte(mm, dst_vma, dst_addr, src_addr, dst_pte, src_pte,
> > > > >                               orig_dst_pte, orig_src_pte, dst_pmd, dst_pmdval,
> > > > >                               dst_ptl, src_ptl, src_folio);
> > > > >
> > > >
> >

Thanks
Barry