On Sat, Apr 07, 2018 at 08:16:46PM -0700, Matthew Wilcox wrote:
> On Fri, Apr 06, 2018 at 03:33:36PM +0200, Dmitry Vyukov wrote:
> > On Fri, Apr 6, 2018 at 3:24 PM, syzbot
> > <syzbot+75397ee3df5c70164154@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > cache_from_obj: Wrong slab cache. names_cache but object is from kmalloc-96
> >
> > Al, do you see how this can happen?
>
> I don't see how it happened, but when looking at this bug, I thought
> "This is very complicated, I think there's a simpler way to handle this".
>
> Here's a proposal. It won't apply to any existing tree (depends on a
> couple of local commits), but I think you'll get the general flavour
> of it. It's mostly compile-tested (build still running, but fs/ and
> kernel/ compiled without issue).
> +struct audit_names;
> +
> +struct filename {
> + const char *name; /* pointer to actual string */
> + const __user char *uptr; /* original userland pointer */
> + struct audit_names *aname;
> +};
>
> /*
> * Each physical page in the system has a struct page associated with
> @@ -188,6 +195,7 @@ struct page {
> spinlock_t ptl;
> #endif
> };
> + struct filename filename;
> };
Oh, lovely - extra 24 bytes into each struct page. Plus the delta to
performance due to switching from kmem_cache_alloc to alloc_page.
Negative one, that is...
