Re: Potentially missing "memory" clobbers in bitops.h for x86

[Peter Zijlstra]

On Fri, Mar 29, 2019 at 03:05:54PM -0700, Paul E. McKenney wrote:
> On Fri, Mar 29, 2019 at 02:51:26PM -0700, H. Peter Anvin wrote:
> > On 3/29/19 2:09 PM, Paul E. McKenney wrote:
> > >>
> > >> Note: the atomic versions of these functions obviously need to have
> > >> "volatile" and the clobber anyway, as they are by definition barriers
> > >> and moving memory operations around them would be a very serious error.
> > > 
> > > The atomic functions that return void don't need to order anything except
> > > the input and output arguments.  The oddness with clear_bit() is that the
> > > memory changed isn't necessarily the quantity referenced by the argument,
> > > if the number of bits specified is large.
> > > 
> > > So (for example) atomic_inc() does not need a "memory" clobber, right?

Correct, and many implementations do not, including x86:

static __always_inline void arch_atomic_inc(atomic_t *v)
{
	asm volatile(LOCK_PREFIX "incl %0"
		     : "+m" (v->counter));
}

> > I don't believe that is true: the code calling it has a reasonable
> > expectation that previous memory operations have finished and later
> > memory operations have not started from the point of view of another
> > processor. You are more of an expert on memory ordering than I am, but
> > I'm 89% sure that there is plenty of code in the kernel which makes that
> > assumption.
> 
> From Documentation/core-api/atomic_ops.rst:

We should delete that file.