Re: [PATCH 2/2] rcu: Keep gpnum and completed fields synchronized

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

On Sat, Dec 11, 2010 at 02:21:04AM +0100, Frederic Weisbecker wrote:
> On Fri, Dec 10, 2010 at 04:58:27PM -0800, Paul E. McKenney wrote:
> > On Sat, Dec 11, 2010 at 01:15:17AM +0100, Frederic Weisbecker wrote:
> > > On Fri, Dec 10, 2010 at 04:04:51PM -0800, Paul E. McKenney wrote:
> > > > On Sat, Dec 11, 2010 at 12:47:11AM +0100, Frederic Weisbecker wrote:
> > > > > On Fri, Dec 10, 2010 at 03:39:20PM -0800, Paul E. McKenney wrote:
> > > > > > On Fri, Dec 10, 2010 at 03:02:00PM -0800, Paul E. McKenney wrote:
> > > > > > > On Fri, Dec 10, 2010 at 10:11:11PM +0100, Frederic Weisbecker wrote:
> > > > > > > > When a CPU that was in an extended quiescent state wakes
> > > > > > > > up and catches up with grace periods that remote CPUs
> > > > > > > > completed on its behalf, we update the completed field
> > > > > > > > but not the gpnum that keeps a stale value of a backward
> > > > > > > > grace period ID.
> > > > > > > > 
> > > > > > > > Later, note_new_gpnum() will interpret the shift between
> > > > > > > > the local CPU and the node grace period ID as some new grace
> > > > > > > > period to handle and will then start to hunt quiescent state.
> > > > > > > > 
> > > > > > > > But if every grace periods have already been completed, this
> > > > > > > > interpretation becomes broken. And we'll be stuck in clusters
> > > > > > > > of spurious softirqs because rcu_report_qs_rdp() will make
> > > > > > > > this broken state run into infinite loop.
> > > > > > > > 
> > > > > > > > The solution, as suggested by Lai Jiangshan, is to ensure that
> > > > > > > > the gpnum and completed fields are well synchronized when we catch
> > > > > > > > up with completed grace periods on their behalf by other cpus.
> > > > > > > > This way we won't start noting spurious new grace periods.
> > > > > > > 
> > > > > > > Also good, queued!
> > > > > > > 
> > > > > > > One issue -- this approach is vulnerable to overflow.  I therefore
> > > > > > > followed up with a patch that changes the condition to
> > > > > > > 
> > > > > > > 	if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
> > > > > > 
> > > > > > And here is the follow-up patch, FWIW.
> > > > > > 
> > > > > > 							Thanx, Paul
> > > > > 
> > > > > Hmm, it doesn't apply on top of my two patches. It seems you have
> > > > > kept my two previous patches, which makes it fail as it lacks them
> > > > > as a base.
> > > > > 
> > > > > Did you intend to keep them? I hope they are quite useless now, otherwise
> > > > > it means there is other cases I forgot.
> > > > 
> > > > One is indeed useless, while the other is useful in combinations of
> > > > dyntick-idle and force_quiescent_state().
> > > 
> > > I don't see how.
> > > 
> > > Before we call __note_new_gpnum(), we always have the opportunity
> > > to resync gpnum and completed as  __rcu_process_gp_end() is called
> > > before.
> > > 
> > > Am I missing something?
> > 
> > If the CPU is already aware of the end of the previous grace period,
> > then __rcu_process_gp_end() will return without doing anything.  But if
> > force_quiescent_state() already took care of this CPU, there is no point
> > in its looking for another quiescent state.  This can happen as follows:
> > 
> > o	CPU 0 notes the end of the previous grace period and then
> > 	enters dyntick-idle mode.
> > 
> > o	CPU 2 enters a very long RCU read-side critical section.
> > 
> > o	CPU 1 starts a new grace period.
> > 
> > o	CPU 0 does not check in because it is in dyntick-idle mode.
> > 
> > o	CPU 1 eventually calls force_quiescent_state() a few times,
> > 	and sees that CPU 0 is in dyntick-idle mode, so tells RCU
> > 	that CPU 0 is in an extended quiescent state.  But the
> > 	grace period cannot end because CPU 2 is still in its
> > 	RCU read-side critical section.
> > 
> > o	CPU 0 comes out of dyntick-idle mode, and sees the new
> > 	grace period.  The old code would nevertheless look for
> > 	a quiescent state, and the new code would avoid doing so.
> > 
> > Unless I am missing something, of course...
> > 
> > 							Thanx, Paul
> 
> Aah, so in your scenario, CPU 0, 1 et 2 are the same node (rnp),
> we have not updated rnp->completed because we still wait for CPU 2.
> 
> Then __rcu_process_gp_end() won't increase the gpnum either
> because rnp->completed is still equal to rdp->completed.
> 
> And later on we call note_new_gpnum() that thinks it has a new
> gp to handle but it's wrong.
> 
> Hence the need to look at the mask level there.

CPUs 0, 1, et 2 are not necessarily on the same node, but other than
that, you have it exactly.  The trick is that force_quiescent_state()
takes global action, so CPU 1 does not need to be on the same node as CPU 0.
Furthermore, an RCU read-side critical section anywhere in the system
will prevent any subsequent grace period from completing, so CPU 2
can be on yet another node.

> This makes all sense!

Thank you for the confirmation, will try testing it out more thoroughly.

							Thanx, Paul

> Thanks!
> 
> > 
> > > Thanks.
> > > 
> > > >  I rebased your earlier two
> > > > out and reworked mine, please see below.  Work better?
> > > > 
> > > > 							Thanx, Paul
> > > > 
> > > > ------------------------------------------------------------------------
> > > > 
> > > > commit c808bedd1b1d7c720546a6682fca44c66703af4e
> > > > Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > > > Date:   Fri Dec 10 15:02:47 2010 -0800
> > > > 
> > > >     rcu: fine-tune grace-period begin/end checks
> > > >     
> > > >     Use the CPU's bit in rnp->qsmask to determine whether or not the CPU
> > > >     should try to report a quiescent state.  Handle overflow in the check
> > > >     for rdp->gpnum having fallen behind.
> > > >     
> > > >     Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > > > 
> > > > diff --git a/kernel/rcutree.c b/kernel/rcutree.c
> > > > index 368be76..530cdcd 100644
> > > > --- a/kernel/rcutree.c
> > > > +++ b/kernel/rcutree.c
> > > > @@ -616,9 +616,17 @@ static void __init check_cpu_stall_init(void)
> > > >  static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
> > > >  {
> > > >  	if (rdp->gpnum != rnp->gpnum) {
> > > > -		rdp->qs_pending = 1;
> > > > -		rdp->passed_quiesc = 0;
> > > > +		/*
> > > > +		 * If the current grace period is waiting for this CPU,
> > > > +		 * set up to detect a quiescent state, otherwise don't
> > > > +		 * go looking for one.
> > > > +		 */
> > > >  		rdp->gpnum = rnp->gpnum;
> > > > +		if (rnp->qsmask & rdp->grpmask) {
> > > > +			rdp->qs_pending = 1;
> > > > +			rdp->passed_quiesc = 0;
> > > > +		} else
> > > > +			rdp->qs_pending = 0;
> > > >  	}
> > > >  }
> > > >  
> > > > @@ -680,19 +688,20 @@ __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_dat
> > > >  
> > > >  		/*
> > > >  		 * If we were in an extended quiescent state, we may have
> > > > -		 * missed some grace periods that others CPUs took care on
> > > > +		 * missed some grace periods that others CPUs handled on
> > > >  		 * our behalf. Catch up with this state to avoid noting
> > > > -		 * spurious new grace periods.
> > > > +		 * spurious new grace periods.  If another grace period
> > > > +		 * has started, then rnp->gpnum will have advanced, so
> > > > +		 * we will detect this later on.
> > > >  		 */
> > > > -		if (rdp->completed > rdp->gpnum)
> > > > +		if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
> > > >  			rdp->gpnum = rdp->completed;
> > > >  
> > > >  		/*
> > > > -		 * If another CPU handled our extended quiescent states and
> > > > -		 * we have no more grace period to complete yet, then stop
> > > > -		 * chasing quiescent states.
> > > > +		 * If RCU does not need a quiescent state from this CPU,
> > > > +		 * then make sure that this CPU doesn't go looking for one.
> > > >  		 */
> > > > -		if (rdp->completed == rnp->gpnum)
> > > > +		if (rnp->qsmask & rdp->grpmask)
> > > >  			rdp->qs_pending = 0;
> > > >  	}
> > > >  }