Re: [PATCH 11/41] nohz/cpuset: Don't turn off the tick if rcu needs it

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

On Thu, May 31, 2012 at 06:01:21PM +0200, Frederic Weisbecker wrote:
> On Wed, May 23, 2012 at 09:27:39AM -0700, Paul E. McKenney wrote:
> > On Wed, May 23, 2012 at 06:06:33PM +0200, Frederic Weisbecker wrote:
> > > On Wed, May 23, 2012 at 08:15:42AM -0700, Paul E. McKenney wrote:
> > > > On Wed, May 23, 2012 at 03:52:09PM +0200, Frederic Weisbecker wrote:
> > > > > > > +#ifdef CONFIG_CPUSETS_NO_HZ
> > > > > > > +static bool can_stop_adaptive_tick(void)
> > > > > > > +{
> > > > > > > +	if (!sched_can_stop_tick())
> > > > > > > +		return false;
> > > > > > > +
> > > > > > > +	/* Is there a grace period to complete ? */
> > > > > > > +	if (rcu_pending(smp_processor_id()))
> > > > > > 
> > > > > > You lost me on this one.  Why can't this be rcu_needs_cpu()?
> > > > > 
> > > > > We already have an rcu_needs_cpu() check in tick_nohz_stop_sched_tick()
> > > > > that prevents the tick to shut down if the CPU has local callbacks to handle.
> > > > > 
> > > > > The rcu_pending() check is there in case some other CPU is waiting for the
> > > > > current one to help completing a grace period, by reporting a quiescent state
> > > > > for example. This happens because we may stop the tick in the kernel, not only
> > > > > userspace. And if we are in the kernel, we still need to be part of the global
> > > > > state machine.
> > > > 
> > > > Ah!  But RCU will notice that the CPU is in dyntick-idle mode, and will
> > > > therefore take any needed quiescent-state action on that CPU's behalf.
> > > > So there should be no need to call rcu_pending() anywhere outside of the
> > > > RCU core code.
> > > 
> > > No. If the tick is stopped and we are in the kernel, we may be using RCU
> > > anytime, so we need to be part of the RCU core.
> > 
> > OK, so the only problem is if we spend a long time CPU-bound in the kernel,
> > where "long" is milliseconds or tens of milliseconds.  In that case, the
> > RCU core will notice that the CPU has not responded but is not idle, for
> > example, in rcu_implicit_dynticks_qs().  It can take action at this point
> > to get the offending CPU to pay attention to RCU.
> > 
> > Does this make sense, or am I still missing something?
> 
> Yeah that's exactly the purpose of the rcu_pending() check before shutting down
> the tick and the IPI to wake it up.

Hmmm...  We appear to be talking past each other.

If you use rcu_pending(), you defeat CONFIG_RCU_FAST_NO_HZ and thus fail
to shut of the tick in situations where the application does a system
call involving an RCU update every few tens of milliseconds.  This is not
good.

What we should do instead is to call rcu_needs_cpu() instead of rcu_pending().
In the common case of short system calls, this will allow the tick to be
turned off a higher fraction of the time with no penalty.  In the very
unusual case where a system call runs CPU-bound for tens of milliseconds,
RCU's existing force_quiescent_state() machinery can easily be used to
force the CPU to pay attention to RCU.

Make sense, or am I missing something?

(And yes, the CONFIG_RCU_FAST_NO_HZ heuristics likely need to be adjusted
to better support adaptive ticks -- try less hard to retire callbacks,
for example.)

							Thanx, Paul