[lttng-dev] [PATCH 7/7] Experiment: Add explicit memory barrier in free_completion()

[Mathieu Desnoyers]

On 2023-03-21 06:21, Ondřej Surý wrote:
>> On 20. 3. 2023, at 19:37, Mathieu Desnoyers <mathieu.desnoyers at efficios.com> wrote:
>>
>> On 2023-03-17 17:37, Ondřej Surý via lttng-dev wrote:
>>> FIXME: This is experiment that adds explicit memory barrier in the
>>> free_completion in the workqueue.c, so ThreadSanitizer knows it's ok to
>>> free the resources.
>>> Signed-off-by: Ondřej Surý <ondrej at sury.org>
>>> ---
>>>   src/workqueue.c | 1 +
>>>   1 file changed, 1 insertion(+)
>>> diff --git a/src/workqueue.c b/src/workqueue.c
>>> index 1039d72..f21907f 100644
>>> --- a/src/workqueue.c
>>> +++ b/src/workqueue.c
>>> @@ -377,6 +377,7 @@ void free_completion(struct urcu_ref *ref)
>>>    struct urcu_workqueue_completion *completion;
>>>      completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
>>> + assert(!urcu_ref_get_unless_zero(&completion->ref));
>>
>> Perhaps what we really want here is an ANNOTATE_UNPUBLISH_MEMORY_RANGE() of some sort ?
> 
> I guess?
> 
> My experience with TSAN tells me, that you need some kind of memory barrier when using acquire-release
> semantics and you do:
> 
> if (__atomic_sub_fetch(obj->ref, __ATOMIC_RELEASE) == 0) {
>    /* __ATOMIC_ACQUIRE needed here */
>     free(obj);
> }
> 
> we end up using following code in BIND 9:
> 
> if (__atomic_sub_fetch(obj->ref, __ATOMIC_ACQ_REL) == 0) {
>     free(obj);
> }
> 
> So, I am guessing after the change of uatomic_sub_return() to __ATOMIC_ACQ_REL,
> this patch should no longer be needed.

Actually we want __ATOMIC_SEQ_CST, which is even stronger than ACQ_REL.

Thanks,

Mathieu

> 
> Ondrej
> --
> Ondřej Surý (He/Him)
> ondrej at sury.org
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com