Re: [Valgrind-developers] Annotations for race detection.

[Konstantin S. <kon...@gm...>]

On Tue, Jun 2, 2009 at 11:54 AM, Bart Van Assche
<bar...@gm...> wrote:
> On Tue, Jun 2, 2009 at 9:36 AM, Konstantin Serebryany
> <kon...@gm...> wrote:
>> On Mon, Jun 1, 2009 at 8:19 PM, Bart Van Assche
>>> One additional remark regarding the ANNOTATE_HAPPENS_* macro's: a data
>>> race detection tool has to allocate some memory to keep track of the
>>> inter-thread ordering imposed by these annotations.
>>
>> True.
>>
>>> Since
>>> ANNOTATE_HAPPENS_AFTER may be invoked multiple times with the same
>>> argument, a tool cannot know when it can free the memory allocated to
>>> implement such an annotation. I have added the
>>> ANNOTATE_HAPPENS_AFTER_DONE() annotation in drd.h for this purpose.
>>
>> I would avoid doing this.
>> Once the program frees the memory which was used as an argument to
>> ANNOTATE_HAPPENS_*, the detector may release the resources.
>> (ThreadSanitizer doesn't do this now. I haven't seen this as a
>> problem, but maybe it is...)
>
> I agree that the tool should free the allocated resources at the time
> the client object is freed. But what should a tool do when the
> argument passed to ANNOTATE_HAPPENS_* is not a valid client address,
> such as in unit tests 30 and 31 ?

Hm. Maybe just leak the resources?
I hate to have annotations that do not represent any synchronization
and are need just for bookkeeping.

In ThreadSanitizer I have to cleanup the whole state from time to time
(usually, once per 10-20 minutes).
This is done because of other reasons, but the leaked objects are
freed as well.

--kcc

>
> Bart.
>