Re: [PATCH v3 4/4] rcu: Unlock non-start node only after accessing its gp_seq_needed

[Paul E. McKenney]

On Mon, May 21, 2018 at 09:43:27PM -0700, Joel Fernandes wrote:
> On Mon, May 21, 2018 at 09:16:51PM -0700, Paul E. McKenney wrote:
> > On Mon, May 21, 2018 at 05:28:23PM -0700, Paul E. McKenney wrote:
> > > On Mon, May 21, 2018 at 05:07:34PM -0700, Joel Fernandes wrote:
> > > > On Mon, May 21, 2018 at 04:25:38PM -0700, Paul E. McKenney wrote:
> > > > > On Sun, May 20, 2018 at 09:42:20PM -0700, Joel Fernandes wrote:
> > > > > > We acquire gp_seq_needed locklessly. To be safe, lets do the unlocking
> > > > > > after the access.
> > > > > 
> > > > > Actually, no, we hold rnp_start's ->lock throughout.  And this CPU (or in
> > > > > the case of no-CBs CPUs, this task) is in charge of rdp->gp_seq_needed,
> > > > > so nothing else is accessing it.  Or at least that is the intent.  ;-)
> > > > 
> > > > I was talking about protecting the internal node's rnp->gp_seq_needed, not
> > > > the rnp_start's gp_seq_needed.
> > > 
> > > Ah, good point, I missed the "if" condition.  This can be argued to work,
> > > sort of, given that we still hold the leaf rcu_node structure's lock,
> > > so that there is a limit to how far grace periods can advance.
> > > 
> > > But the code would of course be much cleaner with your change.
> > > 
> > > > We are protecting them in the loop:
> > > > 
> > > > like this:
> > > > for(...)
> > > > 	if (rnp != rnp_start)
> > > > 		raw_spin_lock_rcu_node(rnp);
> > > > 	[...]
> > > > 	// access rnp->gp_seq and rnp->gp_seq_needed
> > > > 	[...]
> > > > 	if (rnp != rnp_start)
> > > > 		raw_spin_unlock_rcu_node(rnp);
> > > > 
> > > > But we don't need to do such protection in unlock_out ? I'm sorry if I'm
> > > > missing something, but I'm wondering if rnp->gp_seq_needed of an internal
> > > > node can be accessed locklessly, then why can't that be done also in the
> > > > funnel locking loop - after all we are holding the rnp_start's lock through
> > > > out right?
> > > 
> > > I was focused on the updates, and missed the rnp->gp_seq_req access in the
> > > "if" statement.  The current code does sort of work, but only assuming
> > > that the compiler doesn't tear the load, and so your change would help.
> > > Could you please resend with your other two updated patches?  It depends
> > > on one of the earlier patches, so does not apply cleanly as-is.  I could
> > > hand-apply it, but that sounds like a good way to make your updated
> > > series fail to apply.  ;-)
> > > 
> > > But could you also make the commit log explicitly call out the "if"
> > > condition as being the offending access?
> > 
> > Never mind, me being stupid.  I need to apply this change to the original
> > commit "rcu: Make rcu_nocb_wait_gp() check if GP already requested", which
> > I have done with this attribution:
> > 
> > [ paulmck: Move lock release past "if" as suggested by Joel Fernandes. ]
> > 
> > I have rebased my stack on top of the updated commit.
> 
> Cool, makes sense. I am assuming this means I don't have to resend this
> patch, if I do let me know :)

No need.

> Either way, once you push your updated tree to kernel.org, I'll double check
> to make sure the change is in :)

Please see 9624746baf6b ("rcu: Make rcu_nocb_wait_gp() check if GP
already requested") on branch rcu/dev.

								Thanx, Paul