Re: [PATCH] hw/display/ati_2d: Fix buffer overflow in ati_2d_blt (CVE-2021-3638)

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

On 9/7/21 7:38 AM, Philippe Mathieu-Daudé wrote:
> On 9/6/21 9:52 PM, BALATON Zoltan wrote:
>> On Mon, 6 Sep 2021, Philippe Mathieu-Daudé wrote:
>>> (Forgot to Cc Alex for eventual reproducer)
>>>
>>> On 9/6/21 6:44 PM, Mauro Matteo Cascella wrote:
>>>> On Mon, Sep 6, 2021 at 5:31 PM Philippe Mathieu-Daudé
>>>> <philmd@redhat.com> wrote:
>>>>>
>>>>> When building QEMU with DEBUG_ATI defined then running with
>>>>> '-device ati-vga,romfile="" -d unimp,guest_errors -trace ati\*'
>>>>> we get:
>>>>>
>>>>>   ati_mm_write 4 0x16c0 DP_CNTL <- 0x1
>>>>>   ati_mm_write 4 0x146c DP_GUI_MASTER_CNTL <- 0x2
>>>>>   ati_mm_write 4 0x16c8 DP_MIX <- 0xff0000
>>>>>   ati_mm_write 4 0x16c4 DP_DATATYPE <- 0x2
>>>>>   ati_mm_write 4 0x224 CRTC_OFFSET <- 0x0
>>>>>   ati_mm_write 4 0x142c DST_PITCH_OFFSET <- 0xfe00000
>>>>>   ati_mm_write 4 0x1420 DST_Y <- 0x3fff
>>>>>   ati_mm_write 4 0x1410 DST_HEIGHT <- 0x3fff
>>>>>   ati_mm_write 4 0x1588 DST_WIDTH_X <- 0x3fff3fff
>>>>>   ati_2d_blt: vram:0x7fff5fa00000 addr:0 ds:0x7fff61273800
>>>>> stride:2560 bpp:32 rop:0xff
>>>>>   ati_2d_blt: 0 0 0, 0 127 0, (0,0) -> (16383,16383) 16383x16383 > ^
>>>>>   ati_2d_blt: pixman_fill(dst:0x7fff5fa00000, stride:254, bpp:8,
>>>>> x:16383, y:16383, w:16383, h:16383, xor:0xff000000)
>>>>>   Thread 3 "qemu-system-i38" received signal SIGSEGV, Segmentation
>>>>> fault.
>>>>>   (gdb) bt
>>>>>   #0  0x00007ffff7f62ce0 in sse2_fill.lto_priv () at
>>>>> /lib64/libpixman-1.so.0
>>>>>   #1  0x00007ffff7f09278 in pixman_fill () at /lib64/libpixman-1.so.0
>>>>>   #2  0x0000555557b5a9af in ati_2d_blt (s=0x631000028800) at
>>>>> hw/display/ati_2d.c:196
>>>>>   #3  0x0000555557b4b5a2 in ati_mm_write (opaque=0x631000028800,
>>>>> addr=5512, data=1073692671, size=4) at hw/display/ati.c:843
>>>>>   #4  0x0000555558b90ec4 in memory_region_write_accessor
>>>>> (mr=0x631000039cc0, addr=5512, ..., size=4, ...) at
>>>>> softmmu/memory.c:492
>>>>>
>>>>> Commit 584acf34cb0 ("ati-vga: Fix reverse bit blts") introduced
>>>>> the local dst_x and dst_y which adjust the (x, y) coordinates
>>>>> depending on the direction in the SRCCOPY ROP3 operation, but
>>>>> forgot to address the same issue for the PATCOPY, BLACKNESS and
>>>>> WHITENESS operations, which also call pixman_fill().
>>>>>
>>>>> Fix that now by using the adjusted coordinates in the pixman_fill
>>>>> call, and update the related debug printf().
>>>>>
>>>>> Reported-by: Qiang Liu <qiangliu@zju.edu.cn>
>>>>> Fixes: 584acf34cb0 ("ati-vga: Fix reverse bit blts")
>>>>> Signed-off-by: Philippe Mathieu-Daudé <philmd@redhat.com>
>>>>> ---
>>>>>  hw/display/ati_2d.c | 6 +++---
>>>>>  1 file changed, 3 insertions(+), 3 deletions(-)

>> I don't think assigning a CVE to a bug that is in an experimental and
>> largely unused part and happens when one enables debug code really worth
>> the hassle, this could be handled as a normal bug. As long as the
> 
> CVE assignment can happens outside of QEMU community, we try to make it
> clear what is the "security boundary" but researchers filling CVEs
> might not understand it well.

BTW see commit b317006a3f1 ("docs/secure-coding-practices: Describe how
to use 'null-co' block driver") which is related to your suggestion.