Re: [Qemu-devel] [PATCH v1 3/5] include/qemu/atomic.h: default to __atomic functions

[Pranith Kumar <bobby.prani@gmail.com>]

On Mon, Apr 4, 2016 at 1:03 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:

>>
>> Quoting http://www.inf.pucrs.br/~flash/progeng2/cppreference/w/cpp/atomic/atomic_thread_fencehtml.html:
>>
>> "Establishes memory synchronization ordering of non-atomic and relaxed
>> atomic accesses"
>
> You can find some information at
> https://bugzilla.redhat.com/show_bug.cgi?id=1142857 and

This bug seems to be out-of-bounds. I get a message saying that the
bug has been restricted for internal development processes.

> https://retrace.fedoraproject.org/faf/problems/670281/.
>
> I've looked at private email from that time and I was pointed to this
> sentence in GCC's manual, which says the opposite:
>
>     "Note that in the C++11 memory model, fences (e.g.,
>     ‘__atomic_thread_fence’) take effect in combination with other
>     atomic operations on specific memory locations (e.g., atomic loads);
>     operations on specific memory locations do not necessarily affect
>     other operations in the same way."
>
> Basically, the __atomic fences are essentially a way to break up a
> non-relaxed atomic access into a relaxed atomic access and the
> ordering-constraining fence.  According to those emails, atomic
> load-acquires and store-releases can provide synchronization for plain
> loads and stores (not just for relaxed atomic loads and stores).
> However, an atomic thread fence cannot do this.

Yes, on further research... this looks to be the case. I think I
understand the reasoning for this is to separate synchronization
variables/objects vs data.

>
> It depends on the compiler you're using.  With some luck, reverting the
> patch will cause accesses across smp_wmb() or smp_rmb() to get
> reordered.  In our case it was in thread-pool.c; Kevin Wolf did the
> analysis.

If ordering was crucial for those stores, I think it would have been
better to use atomics on them to enforce ordering.

Also, do you plan to introduce load_acquire/store_release() operations
like done in the linux kernel? They would cleanly map to gcc atomics
and make the ordering requirements explicit.

Thanks!
-- 
Pranith