Re: [PATCH v2] drm/i915: bounds check execbuffer relocation count

[Chris Wilson <chris@chris-wilson.co.uk>]

On Mon, Mar 11, 2013 at 03:25:59PM -0700, Kees Cook wrote:
> On Mon, Mar 11, 2013 at 3:00 PM, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> > On Mon, Mar 11, 2013 at 02:23:29PM -0700, Kees Cook wrote:
> >> It is possible to wrap the counter used to allocate the buffer for
> >> relocation copies. This could lead to heap writing overflows.
> >
> > I'd keep the return value as EINVAL so that we can continue to
> > distinguish between the user passing garbage and hitting an oom. And
> > total_relocs is preferrable to total, which also leads us to think more
> > carefully about the error condition. I think the check should be against
> > INT_MAX / sizeof(struct reloc_entry) for consistency with our other
> > guard against overflows whilst allocating.
> 
> I've ended up with this:
> 
>         int max_alloc = INT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
> ...
>                 /* First check for malicious input causing overflow */
>                 if (exec[i].relocation_count > max_alloc)
>                         return -EINVAL;
>                 if (exec[i].relocation_count > max_alloc - total_relocs)
>                         return -EINVAL;
>                 total_relocs += exec[i].relocation_count;
> 
> And looking at that, I wonder if we should just eliminate the first if entirely?

Aye, seems reasonable. So perhaps,

/* First check for malicious input causing overflow in the worst case
 * where we need to allocate the entire relocation tree as a single
 * array.
 */
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre