[Mesa-dev] [RFC PATCH] mesa: Delete VAO _MaxElement code and index buffer bounds checking.

Tue Sep 16 10:04:11 PDT 2014

Am 16.09.2014 17:48, schrieb Brian Paul:
> On 09/15/2014 06:00 PM, Roland Scheidegger wrote:
>> Am 15.09.2014 08:31, schrieb Kenneth Graunke:
>>> Fredrik's implementation of ARB_vertex_attrib_binding introduced new
>>> gl_vertex_attrib_array and gl_vertex_buffer_binding structures, and
>>> converted Mesa's older gl_client_array to be derived state.  Ultimately,
>>> we'd like to drop gl_client_array and use those structures directly.
>>>
>>> One hitch is that gl_client_array::_MaxElement doesn't correspond to
>>> either structure (unlike every other field), so we'd have to figure out
>>> where to store it.  The _MaxElement computation uses values from both
>>> structures, so it doesn't really belong in either place.  We could put
>>> it in the VAO, but we'd have to pass it around everywhere.
>>>
>>> It turns out that it's only used when ctx->Const.CheckArrayBounds is
>>> set, which is only set by the (rarely used) classic swrast driver.
>>> It appears that drivers/x11 used to set it as well, which was intended
>>> to avoid segmentation faults on out-of-bounds memory access in the X
>>> server (probably for indirect GLX clients).  However, ajax deleted that
>>> code in 2010 (commit 1ccef926be46dce3b6b5c76e812e2fae4e205ce7).
>>>
>>> The bounds checking apparently doesn't actually work, either.  Non-VBO
>>> attributes arbitrarily set _MaxElement to 2 * 1000 * 1000 * 1000.
>>> vbo_save_draw and vbo_exec_draw remark /* ??? */ when setting it, and
>>> the i965 code contains a comment noting that _MaxElement is often bogus.
>> Well there's not much you can do for the non-vbo case, since you simply
>> don't know how large that buffer pointed to by that client pointer you
>> were given by the app is...
>>
>>>
>>> Given that the code is complex, rarely used, and dubiously functional,
>>> it doesn't seem worth maintaining going forward.  This patch drops it.
>>>
>>> This will probably mean the classic swrast driver may begin crashing on
>>> out of bounds vertex buffer access in some cases, but I believe that is
>>> allowed by OpenGL (and probably happened for non-VBO accesses anyway).
>>> There do not appear to be any Piglit regressions, either.
>>>
>>> Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
>>> ---
>>>   src/mesa/drivers/dri/i965/brw_draw_upload.c |  5 +--
>>>   src/mesa/drivers/dri/swrast/swrast.c        |  3 --
>>>   src/mesa/main/api_validate.c                | 66
>>> -----------------------------
>>>   src/mesa/main/arrayobj.c                    | 46 --------------------
>>>   src/mesa/main/arrayobj.h                    |  4 --
>>>   src/mesa/main/attrib.c                      |  1 -
>>>   src/mesa/main/context.c                     |  3 --
>>>   src/mesa/main/mtypes.h                      | 10 -----
>>>   src/mesa/main/state.c                       |  5 ---
>>>   src/mesa/main/varray.c                      |  9 +---
>>>   src/mesa/main/varray.h                      | 33 ---------------
>>>   src/mesa/vbo/vbo_exec_array.c               | 26 +++---------
>>>   src/mesa/vbo/vbo_exec_draw.c                |  2 -
>>>   src/mesa/vbo/vbo_save_draw.c                |  1 -
>>>   src/mesa/vbo/vbo_split_copy.c               |  1 -
>>>   15 files changed, 9 insertions(+), 206 deletions(-)
>>>
>>> Hi Brian, Roland, Eric, all...
>>>
>>> What do you think of this idea?  Good idea?  Terrible idea?
>>>
>>> In theory, this seems like a reasonable idea for software drivers, but
>>> it doesn't appear that softpipe/llvmpipe use this code, so I'm not sure
>>> if it's worth maintaining just for swrast...
>> draw does its own validation (according to more strict d3d10 rules,
>> even), as long as everything comes in nice buffers with known sizes this
>> isn't much of a problem (and it doesn't need to be done as a separate
>> pass).
>>
>> I am pretty sure though any app used to be able to crash the X server
>> (when using indirect rendering) when not doing bounds checking pretty
>> easily, but maybe that vanished at some point somehow (luckily, I never
>> had to look at indirect rendering for years...).
>>
>> I think at some point it was also useful for debugging (so you could
>> more easily see where that weird segfault was coming from) though again
>> of course it did nothing for non-vbo arrays. classic swrast could
>> probably reimplement this on its own if nothing else uses it anymore. Of
>> course real hw nowadays you just give the buffer sizes and the hw will
>> make sure nothing is fetched outside bounds on its own.
>>
>> So, for me dropping this looks ok, but I'm not really working much in
>> that area nowadays.
> 
> As Eric pointed out, it doesn't look like there's any glDrawElements or
> VBO code getting used in the X server (the glDrawElements call gets
> "unwound" in the client-side GLX code, AFAICT).
> 
> As it is, the CheckArrayBounds code checks for invalid array indexes at
> draw-validation time.  If a bad index is found, the whole draw is
> discarded.  Alternately, a driver/device can discard individual
> primitives while drawing if an array index is out of bounds (if the
> vertex fetch fails, skip the prim).  The TnL (src/tnl/) module has never
> had any support for per-vertex bounds checking but I believe all the
> gallium drivers (should) handle it.  I think this behavior is what's
> expected of OpenGL/Direct3D nowadays so the CheckArrayBounds approach
> isn't very good.
I believe OpenGL, at least without ARB_robustness, would still allow
dropping the whole draw call, though it's definitely a no go for d3d10.

> 
> In piglit we have tests/spec/arb_robustness/draw-vbo-bounds.c.  It's not
> included in all.py though so it's probably seldom run.  I tested it with
> the software drivers (non-GLX mode):
> 
> swrast: failed assertion, implementation error: bad datatype 0x1010101
> in interpolate_int_colors, plus valgrind errors in the VBO code.  But if
> I enable ctx->Const.CheckArrayBounds, I get warnings but no crashes.
Well if you enable bounds checking which drops the whole draw call, the
test most likely never draws anything (unless you enable the valid
indices only bit).

> 
> llvmpipe: failed assertion, lp_setup_tri.c:495:do_triangle_ccw
Oh interesting. The code does not just test invalid elements, it also
has random vertices. This crash sounds more like being due to the
latter, though clipping should have prevented it (the fixed point dcdx
value didn't fit apparently but we've got 64bit values there so we
should be able to handle any valid vertex with ease). In any case sounds
like something which should be fixed (the assertion in this place is
rather harmless, though the rendering will probably be quite random if
that happens...).

> 
> softpipe: pass
> 
> I just tried this piglit test in Ubuntu 14 with LIBGL_ALWAYS_INDIRECT
> and I get a glx protocol error (no X crash).  But that's using the
> llvmpipe driver.
Ok I'm convinced you can't crash the server like that nowadays :-).

Roland


> 
> Anyway, I think this code can be removed.  Ideally, some index checking
> would be added to the tnl code to try to avoid crashes, but I doubt
> anyone wants to bother with it.
> 
> Reviewed-by: Brian Paul <brianp at vmware.com>
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://urldefense.proofpoint.com/v1/url?u=http://lists.freedesktop.org/mailman/listinfo/mesa-dev&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=F4msKE2WxRzA%2BwN%2B25muztFm5TSPwE8HKJfWfR2NgfY%3D%0A&m=ncC7OVis72yHl9E0MEUdD58NGXXOJnYzxzEpXp4mCvQ%3D%0A&s=5e6141866f7db7315bc080b5ac87359c564a689fad564611baad6b22e9de0746
>