Re: [PATCH] mm: zswap: fix potential memory corruption on duplicate store

[Domenico Cerasuolo]

On Fri, Sep 22, 2023 at 7:42 PM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>
> On Fri, Sep 22, 2023 at 07:22:11PM +0200, Domenico Cerasuolo wrote:
> > While stress-testing zswap a memory corruption was happening when writing
> > back pages. __frontswap_store used to check for duplicate entries before
> > attempting to store a page in zswap, this was because if the store fails
> > the old entry isn't removed from the tree. This change removes duplicate
> > entries in zswap_store before the actual attempt.
> >
> > Based on commit ce9ecca0238b ("Linux 6.6-rc2")
> >
> > Fixes: 42c06a0e8ebe ("mm: kill frontswap")
> > Signed-off-by: Domenico Cerasuolo <cerasuolodomenico@xxxxxxxxx>
>
> Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>
>
> > @@ -1218,6 +1218,19 @@ bool zswap_store(struct folio *folio)
> >       if (!zswap_enabled || !tree)
> >               return false;
> >
> > +     /*
> > +      * If this is a duplicate, it must be removed before attempting to store
> > +      * it, otherwise, if the store fails the old page won't be removed from
> > +      * the tree, and it might be written back overriding the new data.
> > +      */
> > +     spin_lock(&tree->lock);
> > +     dupentry = zswap_rb_search(&tree->rbroot, offset);
> > +     if (dupentry) {
> > +             zswap_duplicate_entry++;
> > +             zswap_invalidate_entry(tree, dupentry);
> > +     }
> > +     spin_unlock(&tree->lock);
>
> Do we still need the dupe handling at the end of the function then?
>
> The dupe store happens because a page that's already in swapcache has
> changed and we're trying to swap_writepage() it again with new data.
>
> But the page is locked at this point, pinning the swap entry. So even
> after the tree lock is dropped I don't see how *another* store to the
> tree at this offset could occur while we're compressing.

My reasoning here was that frontswap used to invalidate a dupe right before
calling store(), so I thought that the check at the end of zswap_store() was
actually needed.
Since we acquire the tree lock anyway to add the new entry to the LRU, wouldn't
checking the result of zswap_rb_insert be a very cheap sanity check? We could
treat it as an error and fail the store(), or just add a warning and still
invalidate the dupe?