On 6 June 2018 at 12:26, Christoph Hellwig <hch@xxxxxx> wrote:
> On Tue, Jun 05, 2018 at 02:10:24PM +0200, David Sterba wrote:
>> On Mon, Jun 04, 2018 at 09:31:16PM +0200, Andreas Gruenbacher wrote:
>> > @@ -1062,8 +1063,9 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
>> > if (ret < 0)
>> > iomap_dio_set_error(dio, ret);
>> >
>> > + smp_mb__before_atomic();
>> > if (!atomic_dec_and_test(&dio->ref)) {
>>
>> The barrier should be documented. I tried to do a quick look around the
>> code if it's clear why it's there but it's not. Thanks.
>
> It doesn't really make sense. smp_mb__before_atomic is for relaxed
> atomic operations, which atomic_dec_and_test is not.
After re-reading Documentation/core-api/atomic_ops.rst again, I agree
it's not needed. There is an example in the documentation that was
confusing me.
Thanks,
Andreas
