Moan: usage of __iormb() and __iowmb() outside of asm/io.h

[will.deacon@arm.com (Will Deacon)]

Hi Russell,

On Mon, Jun 08, 2015 at 07:47:01PM +0100, Russell King - ARM Linux wrote:
> I notice that we have users of __iormb() and __iowmb() outside of asm/io.h:
> 
> drivers/clocksource/timer-keystone.c:   __iowmb();
> drivers/dma/cppi41.c:                   __iormb();
> drivers/dma/cppi41.c:   __iowmb();
> drivers/dma/cppi41.c:           __iowmb();
> drivers/dma/cppi41.c:                   __iormb();
> drivers/soc/ti/knav_qmss_queue.c:       __iowmb();
> 
> These are not official kernel barriers - the only reason they exist in
> asm/io.h is purely to provide a barrier implementation _for_ creating
> the accessors _in_ asm/io.h, which are macros, and therefore these
> macros need to stay around for the same scope as those accessors.
> 
> As with all details which are an architecture matter, they are subject
> to the whims of the architecture maintainer to provide whatever semantics
> for them that the architecture maintainer deems fit: there is no official
> requirement for anything of that nature to do anything, and no guarantee
> that anything such a detail does today it will do so tomorrow.
> 
> This is why only official interfaces should be used, and if they do not
> satisfy the requirements, then new official interfaces need to be 
> proposed.  Don't ever poke about with stuff that's an architecture
> implementation detail.
> 
> We've been here before with some of the cache flushing code - and people
> have been burnt by it.
> 
> I do wish that people would see the difference between stuff which is
> implemented to facilitate the implementation of an architecture detail
> vs something which is provided for everyone's use.
> 
> I'm working on a patch which will completely remove these from view.
> I would strongly suggest that these uses are removed from the above
> code as a matter of urgency.

I agree to removing these from view; we already have plenty of barrier
macros and we don't want these to proliferate outside of the arch code.

Any chance you could do a similar change for arm64, please (we have the
same macros there)?

Will