[Mesa-dev] [PATCH] mesa: fixes for DrawRangeElements[BaseVertex] index validation

[Jose Fonseca]

----- Original Message -----
> On 02/04/2012 06:14 AM, Jose Fonseca wrote:
> >
> >
> > ----- Original Message -----
> >> On 01/27/2012 06:00 PM, Roland Scheidegger wrote:
> >>> in check_index_bounds the comparison needs to be "greater equal"
> >>> since
> >>> contrary to the name _MaxElement is the count of the array.
> >>> In vbo_exec_DrawRangeElementsBaseVertex, take into account the
> >>> basevertex.
> >>> As far as I can tell it is completely ok (though maybe stupid) to
> >>> have
> >>> start/end of 100/199, with _MaxElement being 100, if the
> >>> basevertex
> >>> is -100 (since the start/end are prior to adding basevertex). The
> >>> opposite
> >>> is also true (0/99 is not ok if _MaxElement is 100 and and
> >>> basevertex is 100).
> >>> Previously we would have issued a warning in such cases and worse
> >>> probably
> >>> "fixed" end to a bogus value.
> >>> Totally untested, found by code inspection when looking at bug
> >>> 40361 (which
> >>> seems unrelated after all).
> >>> ---
> >>>    src/mesa/main/api_validate.c  |    2 +-
> >>>    src/mesa/vbo/vbo_exec_array.c |   30
> >>>    +++++++++++++++++++-----------
> >>>    2 files changed, 20 insertions(+), 12 deletions(-)
> >>>
> >>> diff --git a/src/mesa/main/api_validate.c
> >>> b/src/mesa/main/api_validate.c
> >>> index b6871d0..1ae491f 100644
> >>> --- a/src/mesa/main/api_validate.c
> >>> +++ b/src/mesa/main/api_validate.c
> >>> @@ -187,7 +187,7 @@ check_index_bounds(struct gl_context *ctx,
> >>> GLsizei count, GLenum type,
> >>>       vbo_get_minmax_indices(ctx,&prim,&ib,&min,&max, 1);
> >>>
> >>>       if ((int)(min + basevertex)<   0 ||
> >>> -       max + basevertex>   ctx->Array.ArrayObj->_MaxElement) {
> >>> +       max + basevertex>= ctx->Array.ArrayObj->_MaxElement) {
> >>>          /* the max element is out of bounds of one or more
> >>>          enabled
> >>>          arrays */
> >>>          _mesa_warning(ctx, "glDrawElements() index=%u is out of
> >>>          bounds (max=%u)",
> >>>                        max, ctx->Array.ArrayObj->_MaxElement);
> >>
> >> I might split this hunk out into a separate patch.  This hunk is:
> >> Reviewed-by: Kenneth Graunke<kenneth at whitecape.org>
> >>
> >> And yeah, _MaxElement does make me think "index of the last
> >> element"
> >> (N-1), not "the total number of elements" (N).  I'm not a fan of
> >> the
> >> name.
> 
> If someone wants to rename _MaxElement or change its definition,
> that's OK with me.
> 
> 
> [...]
> 
> >> I don't think we want to include basevertex when adjusting 'start'
> >> and
> >> 'end'.  The final call to vbo_validated_drawrangeelements already
> >> passes
> >> basevertex, so I'm pretty sure this will adjust it twice.
> >>
> >> Also, I just sent out a patch (vbo: Ignore invalid element ranges
> >> if
> >> fixing 'end' breaks 'start') and realized it'll conflict here.
> >>
> >> I'm leaning toward dropping this code that clamps "end"
> >> altogether.
> >> It's
> >> awfully dodgy already, and trying to correctly compensate for
> >> basevertex
> >> is pretty much destroying any faith I had left in it working
> >> properly.
> >>
> >> After all, glDrawRangeElementsBaseVertex is just an optimized
> >> version
> >> of
> >> glDrawElementsBaseVertex, so it should be totally safe to just
> >> drop
> >> the
> >> range entirely.  Considering that we've already proven the range
> >> is
> >> insane, that seems like the safest option.  Plus...optimizing
> >> broken
> >> calls?  Really?
> >
> > I'm OK with passing the ranges unmodified to the drivers. And let
> > them handle them. But the drivers (or the hardware) needs to be
> > robust against fetching vertices beyond the vertex buffer's size.
> > I'm not sure if they all do that.
> 
> In addition to debugging, the other reason I added the _MaxElement
> stuff was to try to catch out-of-bounds accesses and prevent
> segfaults, in particular for when Mesa is running in the X server.
> Otherwise, one could bring down the X server with one trivial
> glDrawElements call.

Just to be clear, I'm not suggesting eliminating out-of-bounds checks -- out-of-bounds should still be prevented IMO --, but simply suggesting to move the checks further down the stack, as I think is much easier to clamp vertex/index fetches when actually fetching them (by comparing memory accesses against the buffer bounds, or by programming those buffer bounds in the hardware), than trying to reverse compute the maximum index that will not cause any buffer overflow. Specially in face of index biases, instancing, ARB_shader_image_load_store, and who knows what else in the future.  I have the strong suspicion that from a mathematical point of view, it is not a well defined problem, and that any formula we come up with will either err on one way or the other.

Jose