Re: + atomic-improve-atomic_inc_unless_negative-atomic_dec_unless_positive .patch added to -mm tree

[Oleg Nesterov <oleg@redhat.com>]

On 03/15, Paul E. McKenney wrote:
>
> On Fri, Mar 15, 2013 at 07:34:32PM +0100, Frederic Weisbecker wrote:
> > 2013/3/15 Oleg Nesterov <oleg@redhat.com>:
> > >
> > > My point was: should we fix atomic_add_unless() then? If not, why
> > > should atomic_add_unless_negative() differ?
> >
> > They shouldn't differ I guess.
>
> Completely agreed.  It is not like memory ordering is simple, so we should
> keep the rules simple.

It is hardly possible to argue with this ;)

> The rule is that if an atomic primitive returns non-void, then there is
> a full memory barrier before and after.

This case is documented...

> This applies to primitives
> returning boolean as well, with atomic_dec_and_test() setting this
> precedent from what I can see.

I don't think this is the "fair" comparison. Unlike atomic_add_unless(),
atomic_dec_and_test() always changes the memory even if it "fails".

If atomic_add_unless() returns 0, nothing was changed and if we add
the barrier it is not clear what it should be paired with.


But OK. I have to agree that "keep the rules simple" makes sense, so
we should change atomic_add_unless() as well.

Oleg.