Re: [RFC PATCH 1/1] kernel/rcu/tree.c: correct a check for grace period in progress

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

On Wed, Jun 11, 2014 at 12:23:57AM -0400, Pranith Kumar wrote:
> Hi Paul,
> 
> On Wed, Jun 11, 2014 at 12:12 AM, Paul E. McKenney
> <paulmck@linux.vnet.ibm.com> wrote:
> >>       if (rnp->gpnum != rnp->completed ||
> >> -         ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
> >> +         ACCESS_ONCE(rnp_root->gpnum) != ACCESS_ONCE(rnp_root->completed)) {
> >
> > At this point in the code, we are checking the current rcu_node structure,
> > which might or might not be the root.  If it is not the root, we absolutely
> > cannot compare against the root because we don't yet hold the root's lock.
> >
> 
> I was a bit thrown by the double checking which is being done
> (rnp->gpnum != rnp->complete) in that if condition. Once without
> ACCESS_ONCE and one with. Is there any particular reason for this?
> 
> I now understand that we are comparing ->gpnum and ->completed of the
> root node which might change from under us if we don't hold the root's
> lock. I will keep looking :)

Hmmm...  Now that you mention it, that does look a bit strange.

							Thanx, Paul