----- On Sep 25, 2017, at 8:25 AM, Peter Zijlstra peterz@xxxxxxxxxxxxx wrote:
> On Mon, Sep 25, 2017 at 08:10:54PM +0800, Boqun Feng wrote:
>> > static void membarrier_register_private_expedited(void)
>> > {
>> > struct task_struct *p = current;
>> >
>> > if (READ_ONCE(p->mm->membarrier_private_expedited))
>> > return;
>> > membarrier_arch_register_private_expedited(p);
>
> Should we not then also do:
>
> barrier();
>
>> > WRITE_ONCE(p->mm->membarrier_private_expedited, 1);
>> > }
>
> to avoid the compiler lifting that store?
membarrier_arch_register_private_expedited() being a function call, I
recall compilers cannot move load/stores across those. Moreover, even if
that function would happen to be eventually inlined, synchronize_sched()
is needed at the end of the function to ensure the scheduler will observe
the thread flags before it returns. That too would then act as a compiler
barrier if that function is ever inlined in the future.
So do you think we should still add the barrier() as documentation, or is
having synchronize_sched() in the callee enough ?
By the way, I think I should add a READ_ONCE() in membarrier_private_expedited
to pair with the WRITE_ONCE() in registration, such as:
if (!READ_ONCE(current->mm->membarrier_private_expedited))
return -EPERM;
Thanks!
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
