On Thu, May 30, 2019 at 12:34:05PM +0200, Andrea Parri wrote:
> On Wed, May 29, 2019 at 07:38:46PM +0200, Jann Horn wrote:
> > On Wed, May 29, 2019 at 6:21 PM Oleg Nesterov <o...@redhat.com> wrote:
> > > (I am wondering if smp_acquire__after_ctrl_dep() could be used instead,
> > > just to
> > > make this code look more confusing)
> >
> > Uuh, I had no idea that that barrier type exists. The helper isn't
> > even explicitly mentioned in Documentation/memory-barriers.rst. I
> > don't really want to use dark magic in the middle of ptrace access
> > logic...
Yeah, it's sorta not documented on purpose. It's too easy to get wrong
and we've only used it inside a number of more convenient primitives as
an optimzation.
I suppose we could add it to the section on control dependencies; just
to scare more people :-)
> > Anyway, looking at it, I think smp_acquire__after_ctrl_dep() doesn't
> > make sense here; quoting the documentation: "A load-load control
> > dependency requires a full read memory barrier, not simply a data
> > dependency barrier to make it work correctly". IIUC
> > smp_acquire__after_ctrl_dep() is for cases in which you would
> > otherwise need a full memory barrier - smp_mb() - and you want to be
> > able to reduce it to a read barrier.
>
> It is supposed to be used when you want an ACQUIRE but you only have a
> control dependency (so you "augment the dependency" with this barrier).
>
> FWIW, I do agree on the "dark magic"..., and I'd strongly recommend to
> not use this barrier (or, at least, to use it with high suspicion).
Right, so the purpose of the barrier is to upgrade a LOAD->STORE order
(as provided by the ctrl-dep) to a LOAD->{LOAD,STORE} order as would be
provided by load-acquire.
