Re: [PATCH] atomic: improve atomic_inc_unless_negative/atomic_dec_unless_positive

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Tue, Mar 12, 2013 at 12:03:23PM +0800, Ming Lei wrote:
> On Tue, Mar 12, 2013 at 11:39 AM, Paul E. McKenney
> <paulmck@linux.vnet.ibm.com> wrote:
> >
> > Atomic operations that return a value are required to act as full memory
> > barriers.  This means that code relying on ordering provided by these
> > atomic operations must also do ordering, either by using an explicit
> > memory barrier or by relying on guarantees from atomic operations.
> >
> > For example:
> >
> >         CPU 0                                   CPU 1
> >
> >         X = 1;                                  r1 = Z;
> >         if (atomic_inc_unless_negative(&Y)      smp_mb();
> >                 do_something();
> >         Z = 1;                                  r2 = X;
> >
> > Assuming X and Z are initially zero, if r1==1, we are guaranteed
> > that r2==1.  However, CPU 1 needs its smp_mb() in order to pair with
> > the barrier implicit in atomic_inc_unless_negative().
> >
> > Make sense?
> 
> Yes, it does, and thanks for the explanation.
> 
> But looks the above example is not what Frederic described:
> 
> "the above atomic_read() might return -1 because there is no
> guarantee it's seeing the recent update on the remote CPU."
> 
> Even I am not sure if adding one smp_mb() around atomic_read()
> can guarantee that too.

Frederic was likely thinking of some other scenario that would be
broken by atomic_inc_unless_negative() failing to act as a full
memory barrier.  Here is another example:


        CPU 0                                   CPU 1

        	                                X = 1;
        if (atomic_inc_unless_negative(&Y)      r1 = atomic_xchg(&Y, -1);
                r2 = X;

If atomic_inc_unless_negative() acts as a full memory barrier, then
if CPU 0 reaches the assignment from X, the results will be guaranteed
to be 1.  Otherwise, there is no guarantee.

							Thanx, Paul