Re: [PATCH 1/2] media: v4l2-core: Shifting signed 32-bit value by 31 bits error

[Shuah Khan <skhan@linuxfoundation.org>]

On 6/11/19 4:27 PM, Shuah Khan wrote:
> On 6/11/19 2:50 PM, Hans Verkuil wrote:
>> On 6/11/19 9:42 PM, Shuah Khan wrote:
>>> On 6/6/19 12:33 AM, Hans Verkuil wrote:
>>>> On 6/6/19 5:22 AM, Randy Dunlap wrote:
>>>>> On 6/5/19 2:53 PM, Shuah Khan wrote:
>>>>>> Fix the following cppcheck error:
>>>>>>
>>>>>> Checking drivers/media/v4l2-core/v4l2-ioctl.c ...
>>>>>> [drivers/media/v4l2-core/v4l2-ioctl.c:1370]: (error) Shifting 
>>>>>> signed 32-bit value by 31 bits is undefined behaviour
>>>>>>
>>>>>> Signed-off-by: Shuah Khan <skhan@linuxfoundation.org>
>>>>>> ---
>>>>>>    drivers/media/v4l2-core/v4l2-ioctl.c | 2 +-
>>>>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>
>>>>>> diff --git a/drivers/media/v4l2-core/v4l2-ioctl.c 
>>>>>> b/drivers/media/v4l2-core/v4l2-ioctl.c
>>>>>> index 6859bdac86fe..333e387bafeb 100644
>>>>>> --- a/drivers/media/v4l2-core/v4l2-ioctl.c
>>>>>> +++ b/drivers/media/v4l2-core/v4l2-ioctl.c
>>>>>> @@ -1364,7 +1364,7 @@ static void v4l_fill_fmtdesc(struct 
>>>>>> v4l2_fmtdesc *fmt)
>>>>>>                        (char)((fmt->pixelformat >> 8) & 0x7f),
>>>>>>                        (char)((fmt->pixelformat >> 16) & 0x7f),
>>>>>>                        (char)((fmt->pixelformat >> 24) & 0x7f),
>>>>>> -                    (fmt->pixelformat & (1 << 31)) ? "-BE" : "");
>>>>>> +                    (fmt->pixelformat & BIT(31)) ? "-BE" : "");
>>>>>>                break;
>>>>>>            }
>>>>>>        }
>>>>>>
>>>>>
>>>>> If this builds, I guess #define BIT(x) got pulled in indirectly
>>>>> since bits.h nor bitops.h is currently #included in that source file.
>>>>>
>>>
>>> It does build. You are right that I should have included bitops.h
>>>
>>>>> Documentation/process/submit-checklist.rst rule #1 says:
>>>>> 1) If you use a facility then #include the file that defines/declares
>>>>>      that facility.  Don't depend on other header files pulling in 
>>>>> ones
>>>>>      that you use.
>>>>>
>>>>> Please add #include <linux/bits or bitops.h>
>>>>>
>>>>
>>>> I'm not sure about this patch. '1 << 31' is used all over in the 
>>>> kernel,
>>>> including in public headers (e.g. media.h, videodev2.h).
>>>>
>>>> It seems arbitrary to change it only here, but not anywhere else.
>>>>
>>>
>>> Right. We have several places in the kernel that do that.
>>>
>>>> In this particular example for the fourcc handling I would prefer to 
>>>> just
>>>> use '1U << 31', both in v4l2-ioctl.c and videodev2.h.
>>>>
>>>
>>> If you would like to take the patch, I can send v2 fixing it using
>>> 1U << 31 - This is simpler since it doesn't nee additional includes.
>>
>> I would like to have this cleaned up in the public media APIs. Those 
>> can be
>> used by other compilers as well and it makes sense to me not to have
>> undefined behavior in those headers.
>>
> 
> Great. That is a good point. I will start looking at the public media
> APIs.
> 
>>>
>>>> A separate patch doing the same for MEDIA_ENT_ID_FLAG_NEXT in 
>>>> media.h would
>>>> probably be a good idea either: that way the public API at least 
>>>> will do
>>>> the right thing.
>>>>
> 
> Sounds good.
> 
>>>
>>> I should have explained it better. I wanted to start with one or two
>>> places first to see if it is worth our time to fix these:
>>>
>>> The full kernel cppcheck log for "Shifting signed 32-bit value by 31
>>> bits is undefined behaviour" can be found at:
>>>
>>> https://drive.google.com/file/d/19Xu7UqBGJ7BpzxEp92ZQYb6F8UPrk3z3/view
>>
>> I don't think it makes sense to fix this for drivers. If gcc would do 
>> this
>> wrong, we'd have noticed it ages ago.
> 

Did some research into this. We are fine with gcc, however looks like
clang had the problem which has been fixed very recently in late 2018
in 6.0 release. This will be a problem even for kernel/drivers if older
clang is used to build it.

I am sending the two header fixes (media.h, and videodev2.h) to start
with.

thanks,
-- Shuah