Re: [PATCH tip/core/locking 4/4] Documentation/memory-barriers.txt: Document ACCESS_ONCE()

[Ingo Molnar <mingo@kernel.org>]

* Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:

> > So, what I don't see this statement cover (and I might be dense about 
> > it!) is whether two ACCESS_ONCE() macros referring to different 
> > variables are allowed to be reordered with each other.
> > 
> > If the compiler reorders:
> > 
> > 	ACCESS_ONCE(x);
> > 	ACCESS_ONCE(y);
> > 
> > to:
> > 
> > 	ACCESS_ONCE(y);
> > 	ACCESS_ONCE(x);
> > 
> > then AFAICS it still meets the "compiler need only forget the contents 
> > of the indicated memory located" requirement that you listed, right?
> 
> True, but if the compiler was willing to reorder ACCESS_ONCE()'s 
> volatile accesses, it would be really hard to write reliable device 
> drivers. [...]

But nowhere do we link ACCESS_ONCE() to 'volatile' semantics in the 
document, do we? (and I'm not sure we should.)

[ In theory a future compiler could offer a smarter, more flexible 
  'compiler barrier' implementation - at which point we might be 
  tempted to use that new facility to implement ACCESS_ONCE(). At that 
  point this ambiguity might arise. ]


> [...]  The standard says the following:
> 
> 	Access to volatile objects are evaluated strictly according to
> 	the rules of the abstract machine.
> 
> That said, compiler writers and standards wonks will argue endlessly 
> about exactly what that does and does not mean.  :-/
> 
> I added a sentence reading:
> 
> 	Of course, the compiler must also respect the order in which
> 	the ACCESS_ONCE()s occur, though the CPU of course need not do so.
> 
> To the end of that paragraph.  Does that help?

Yeah, that looks perfect!

Thanks,

	Ingo