Re: [RFC] block: change to use atomic_inc_return_release()

[yalin wang <yalin.wang2010@gmail.com>]

> On Nov 17, 2015, at 11:38, Jens Axboe <axboe@kernel.dk> wrote:
> 
> On 11/16/2015 08:24 PM, yalin wang wrote:
>> Some arch define this atomic_inc_return_release() OP.
> 
> That is a very vague commit message, you'll need a whole lot more than that... A commit message is supposed to describe the reason for the change. You provide no reason for the change.
> 
>> diff --git a/block/bio.c b/block/bio.c
>> index fbc558b..b251857 100644
>> --- a/block/bio.c
>> +++ b/block/bio.c
>> @@ -310,8 +310,7 @@ static void bio_chain_endio(struct bio *bio, int error)
>>  static inline void bio_inc_remaining(struct bio *bio)
>>  {
>>  	bio->bi_flags |= (1 << BIO_CHAIN);
>> -	smp_mb__before_atomic();
>> -	atomic_inc(&bio->__bi_remaining);
>> +	atomic_inc_return_release(&bio->__bi_remaining);
> 
> Are these equivalent? Where's the documentation for this primitive? The previous code ensured that we ordered the dec of the remaining count with the update of the flags.
> 
i just have a look at ARM64 implementation for this new atomic OP ,
but i don’t find doc in memory-barrier.txt . so i make this RFC for some response,
atomic_inc_return_release()  should have store_release() class memory barriers .
in this example,  smp_store_release() memory barrier is not enough ?
just make sure bi_flags update can been seen by other cores before update atomic counter.
atomic_inc_return_{release,acquire,relax} OP seems newly add to kernel .
But i don’t see much users in code .
Can it be used to replace lots of smp_mb__before_atomic() ?

Thanks