[Intel-gfx] [RFC] drm/i915: prevent out of range pt in the PDE macros (take 2)

[Chris Wilson]

On Tue, Jun 16, 2015 at 02:45:39PM +0100, Dave Gordon wrote:
> On 15/06/15 11:53, Chris Wilson wrote:
> > On Mon, Jun 15, 2015 at 11:33:37AM +0100, Dave Gordon wrote:
> >> On 13/06/15 09:28, Chris Wilson wrote:
> >>> On Fri, Jun 12, 2015 at 06:30:56PM -0300, Paulo Zanoni wrote:
> >>>> From: Paulo Zanoni <paulo.r.zanoni at intel.com>
> >>>>
> >>>> We tried to fix this in the following commit:
> >>>>
> >>>> commit fdc454c1484a20e1345cf4e4d7a9feaee814147f
> >>>> Author: Michel Thierry <michel.thierry at intel.com>
> >>>> Date:   Tue Mar 24 15:46:19 2015 +0000
> >>>>     drm/i915: Prevent out of range pt in gen6_for_each_pde
> >>>>
> >>>> but the static analyzer still complains that, just before we break due
> >>>> to "iter < I915_PDES", we do "pt = (pd)->page_table[iter]" with an
> >>>> iter value that is bigger than I915_PDES. Of course, this isn't really
> >>>> a problem since no one uses pt outside the macro. Still, every single
> >>>> new usage of the macro will create a new issue for us to mark as a
> >>>> false possitive.
> >>>>
> >>>> After the commit mentioned above we also created some new versions of
> >>>> the macros, so they carry the same "problem".
> >>>>
> >>>> In order to "solve" this "problem", let's leave the macro with a NULL
> >>>> value for pt. So if somebody uses it, we're more likely to get a big
> >>>> error message instead of some silent failure. I hope the static
> >>>> analyzer won't complain about the new solution (I don't have a way to
> >>>> check this!).
> >>>>
> >>>> I know, the solution looks really ugly. I am hoping the reviewers will
> >>>> help us decide if we prefer this patch or if we prefer to keep marking
> >>>> things as false positives.
> >>>>
> >>>> Cc: Michel Thierry <michel.thierry at intel.com>
> >>>> Signed-off-by: Paulo Zanoni <paulo.r.zanoni at intel.com>
> >>>> ---
> >>>>  drivers/gpu/drm/i915/i915_gem_gtt.h | 13 +++++++++----
> >>>>  1 file changed, 9 insertions(+), 4 deletions(-)
> >>>>
> >>>> I sent this as an RFC because I really don't know if complicating the
> >>>> macro even more will help us in any way. I won't really be surprised
> >>>> if I see NACKs on this patch, so don't hesitate if you want to.
> >>>>
> >>>> Also, all I did was boot a Kernel with this patch and make sure it
> >>>> shows the desktop. So consider this as untested, possibly broken.
> >>>>
> >>>> diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h
> >>>> index 0d46dd2..b202ca0 100644
> >>>> --- a/drivers/gpu/drm/i915/i915_gem_gtt.h
> >>>> +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h
> >>>> @@ -352,7 +352,8 @@ struct i915_hw_ppgtt {
> >>>>   */
> >>>
> >>> Overallocate page_table etc by one and put a NULL sentinel in it.
> >>>
> >>> for ((iter) = gen6_pde_index(start); \
> >>>      (length) > 0 && (pt = (pd)->page_table[iter]); \
> >>>      (iter)++, \
> >>>      temp = ALIGN(start+1, 1 << GEN6_PDE_SHIFT) - start, \
> >>>      temp = min_t(unsigned, temp, length), \
> >>>
> >>> -Chris
> >>
> >> This might trigger different warnings from some static analysers, as
> >> 'pt' doesn't get assigned at all if length == 0.
> > 
> > And? If pt is used when length==0 then I would agree with the analyzer
> > that pt should be invalid. If the analyzer can't tell that length is
> > non-zero in the use case and gives false positives, then the analyzer is
> > likely missing genuinine bugs in other cases.
> > -Chris
> 
> If you overallocate as suggested then you can keep the assignment to
> 'pt' first (i.e. unconditional, before the length test) so even a dumb
> analyser won't get confused. OTOH, page_table[] is currently an array of
> 512 pointers which is (or can be) nicely page-aligned, whereas
> increasing it to 513 will make them not fit so nicely :(

Good point.

> Perhaps the simplest way to write the test is:
> 
>     for ((iter) = gen6_pde_index(start);             \
>          (pt) = (length) > 0 && (iter) < I915_PDES ? \
>                   (pd)->page_table[iter] : NULL;     \
>          (iter)++, ...
> 
> which always assigns 'pt', and always leaves it NULL on loop exit.

Just one level of parenthesis required to shut gcc up, but that indeed
does look the neatest way of writing it so far.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre