Re: [PATCH] splice : two smp_mb() can be omitted

[Nick Piggin <nickpiggin@yahoo.com.au>]

Eric Dumazet wrote:

> Nick Piggin a écrit :
>
>>
>> Again, lock / unlock operations require acquire / release 
>> consistency. This is a
>> memory ordering operation. It is not equivalent to smp_mb, though.
>
>
> This thread just show how difficult it is to have consistent use of 
> all this stuff in all kernel. Maybe it is just me ? Should I work on 
> IA64 to have a chance to learn ?


No need, just don't go thinking that mutex_unlock implies smp_mb.

spin_unlock has never implied an smp_rmb on i386.

> For example, Documentation/atomic_ops.txt comments about 
> atomic_inc_return() and atomic_dec_return() seems in contradiction 
> with itself.
>
> --------------------------
>
> Unlike the above routines, it is required that explicit memory
> barriers are performed before and after the operation.  It must be
> done such that all memory operations before and after the atomic
> operation calls are strongly ordered with respect to the atomic
> operation itself.
>
> -------------------------
>
> When I read this, I understand we (the user of such functions) need to 
> add smp_mb(). (That is, those functions wont do it themselves)


This is written from the point of view of the _implementor_. I agree it 
is a bit
confusing, but does the example below clear it up?

>
> Then following text is :
>
> ----------------------------
> For example, it should behave as if a smp_mb() call existed both
> before and after the atomic operation.
>
> --------------------------
>
> Now I understand the reverse.


Now you understand correctly ;)

--

Send instant messages to your online friends http://au.messenger.yahoo.com