[lttng-dev] [PATCH 04/11] urcu/arch/generic: Use atomic builtins if configured

[Mathieu Desnoyers]

On 6/21/23 20:53, Olivier Dion wrote:
> On Wed, 21 Jun 2023, "Paul E. McKenney" <paulmck at kernel.org> wrote:
>> On Mon, May 15, 2023 at 04:17:11PM -0400, Olivier Dion wrote:
>>>   #ifndef cmm_mb
>>>   #define cmm_mb()    __sync_synchronize()
>>
>> Just out of curiosity, why not also implement cmm_mb() in terms of
>> __atomic_thread_fence(__ATOMIC_SEQ_CST)?  (Or is that a later patch?)
> 
> IIRC, Mathieu and I agree that the definition of a thread fence -- acts
> as a synchronization fence between threads -- is too weak for what we
> want here.  For example, with I/O devices.
> 
> Although __sync_synchronize() is probably an alias for a SEQ_CST thread
> fence, its definition -- issues a full memory barrier -- is stronger.
> 
> We do not want to rely on this assumption (alias) and prefer to rely on
> the documented definition instead.
> 

We should document this rationale with a new comment near the #define,
in case anyone mistakenly decides to use a thread fence there to make it
similar to the rest of the code in the future.

Thanks,

Mathieu


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