Re: [Valgrind-users] question about helgrind disable enable checking.

[Rich P. <pro...@gm...>]

Hello,
Suppose that I have a shared variable that is always updates when a
mutex is held and is sometimes read without holding the mutex.  The
read race is fine for my application.  However, I want to use helgrind
to check for other races.

If I disable/enable checking around the lock free read, then I get
spurious write race helgrind reports WRT the locked writes.

So, disable/enable checking is not going to work.

I can add a helgrind suppression for the unlocked read race.  However,
there are write races reported for all the places that the variable is
written.

I could just disable checking for this variable when the variable is
created.  However, that will lose ALL helgrind checking of the
variable.  Is there a better technique?

Thanks

On Sun, Aug 9, 2015 at 10:29 AM, Philippe Waroquiers
<phi...@sk...> wrote:
> On Fri, 2015-08-07 at 10:36 -0400, Rich Prohaska wrote:
>> Hello,
>>
>> I have observed that helgrind will report erroneous data races on a
>> variable if checking on that variable is disable and enabled by a
>> concurrent thread.  Can someone explain what is going on?
> VALGRIND_HG_ENABLE_CHECKING implementation consists in declaring
> that the address range is 'new' and belongs (exclusively) to this
> thread. In other words, calling VALGRIND_HG_ENABLE_CHECKING
> is like if the memory was just allocated.
>
> So, using that definition, it is normal that helgrind reports
> a race, because there is no synchronisation between the thread
> that 'allocated freshly' count, and the thread that does a write
> to count.
>
> Philippe
>
>