Re: [PATCH 3/4] device-core: use atomic_set on .realized property

[Maxim Levitsky]

On Mon, 2020-05-04 at 13:22 +0200, Paolo Bonzini wrote:
> On 04/05/20 12:45, Stefan Hajnoczi wrote:
> > > @@ -983,7 +983,7 @@ static void device_set_realized(Object *obj, bool 
> > > value, Error **errp)
> > >      }
> > >  
> > >      assert(local_err == NULL);
> > > -    dev->realized = value;
> > > +    atomic_set(&dev->realized, value);
> > 
> > A memory barrier is probably needed so that the atomic_read() thread
> > sees up-to-date dev fields.
> 
> Yes, it should be a store-release for the false->true case.  The
> true->false case probably doesn't matter as much.
> 
> Paolo
> 
I was under impression that atomic_set implies a barrier, but now indeed it 
looks like it doesn't.
I''l read upon this a bit and then send an updated patch.


For RCU, sorry for not knowing the details yet, I was under impression that for 
reads you need the rcu read lock
and for writes you also need the RCU read lock, since I first would read then 
write the data, 
plus follow the RCU rule of the update (read, copy, update),
with an atomic swap of a pointer to point to the new copy, and finally register 
a callback with RCU so it frees the old
copy when all the readers of the old copy are guaranteed to be gone.

Best regards,
        Maxim Levitsky