Re: drivers: random: Shift out-of-bounds in _mix_pool_bytes

["H. Peter Anvin" <hpa@zytor.com>]

On 10/24/2014 02:05 PM, One Thousand Gnomes wrote:
> On Fri, 24 Oct 2014 19:10:49 +0400
> Dmitry Vyukov <dvyukov@google.com> wrote:
> 
>> On Fri, Oct 24, 2014 at 7:04 PM, Sasha Levin <sasha.levin@oracle.com> wrote:
>>> On 10/24/2014 09:42 AM, Peter Zijlstra wrote:
>>>> On Fri, Oct 24, 2014 at 09:23:35AM -0400, Sasha Levin wrote:
>>>>>
>>>>> i >> 32 may happen to be "i", but is there anything that prevents the compiler
>>>>> from returning, let's say, 42?
>>>>
>>>> Not really, although gcc seems to opt for the 'sane' option and emit the
>>>> instruction and let the arch figure out how to deal with it. Hence the
>>>> 'fun' difference between x86 and ARM.
>>>
>>> It's interesting how many different views on undefined behaviour there are between
>>> kernel folks.
>>>
>>> Everything between Ted Ts'o saying that GCC can launch nethack on oversized shifts,
>>> to DaveM saying he will file a GCC bug if the behaviour isn't sane w.r.t to memcpy().
>>
>> One of the benefits of fixing such issues (or not letting them into
>> code in the first place) is just saving numerous hours of top-notch
>> engineers spent on disputes like this.
> 
> Also it means when someone quietly changes the default behaviour next
> year in the compiler they won't spend months trying to work out why it
> broke.
> 
> gcc has one behaviour but people also try and build the kernel with icc
> and with llvm. In addition in some cases you risk the compiler simply
> generating an undefined in hardware operation and the hardware behaviour
> changing. If x >> 32 is undefined then generating "load Y with the
> shift, shift X left by Y" is fine. What happens in future silicon - who
> knows.
> 
> Most of the kernel is already very careful about the >> 32 problem.
> 

The real question is if we can rely on the gcc-ism:

	(x >> (S-y)) | (x << y)

... where S is the number of bits to indicate a rotate.

This is technically a gcc extension to the C language.

	-hpa