Re: [PATCH v2 16/18] x86: io: implement dummy relaxed accessor macros for writes

["H. Peter Anvin" <hpa@zytor.com>]

On 05/23/2014 07:57 AM, Will Deacon wrote:
> On Fri, May 23, 2014 at 03:53:20PM +0100, H. Peter Anvin wrote:
>> On 05/23/2014 07:46 AM, Will Deacon wrote:
>>>
>>> I would like the relaxed accessors to be ordered with respect to each other...
>>>
>>> What do you think?
>>>
>>
>> I think "I would like" isn't a very good motivation.  What are the
>> semantics of these things supposed to be?  It seems more than a bit odd
>> to require them to be ordered with respect to each other and everything
>> else (which is what a memory clobber does) and then call them "relaxed".
> 
> I suggested some informal semantics in the cover letter:
> 
>   https://lkml.org/lkml/2014/4/17/269
> 
> Basically, if we define relaxed accesses not to be ordered against anything
> apart from other accesses (relaxed or otherwise) to the same device, then
> they become a tonne cheaper on arm/arm64/powerpc. Currently we have to
> include expensive memory barriers in order to synchronise with accesses to
> DMA buffers which is rarely needed.
> 
> For those requirements, I don't think we need the "memory" clobber for x86,
> but would appreciate your views on this.
> 

OK... first of all you didn't send the cover letter to the union of all
the people you sent patches to, but second, documenting semantics in the
one piece of the patchset that wouldn't make it into git is just about
the worst possible place to put it.

This documentation is absolutely critical if we expect people to be able
to use these correctly, including when additional barriers may be required.

As far as x86 is concerned, in gcc volatiles are ordered with respect to
each other, so as you say I don't think we need a memory clobber here.

	-hpa