Re: [PATCH 2/4] timer: relax tick stop in idle entry

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

On Mon, Nov 16, 2015 at 05:41:03PM -0800, Josh Triplett wrote:
> On Mon, Nov 16, 2015 at 03:26:40PM -0800, Paul E. McKenney wrote:
> > On Mon, Nov 16, 2015 at 02:32:11PM -0800, Josh Triplett wrote:
> > > On Mon, Nov 16, 2015 at 01:51:26PM -0800, Jacob Pan wrote:
> > > > On Mon, 16 Nov 2015 16:06:57 +0100 (CET)
> > > > Thomas Gleixner <tglx@linutronix.de> wrote:
> > > > 
> > > > > >           <idle>-0     [000]    30.093474: bprint:
> > > > > > __tick_nohz_idle_enter: JPAN: tick_nohz_stop_sched_tick 609 delta
> > > > > > 1000000 [JP] but sees delta is exactly 1 tick away. didn't stop
> > > > > > tick.  
> > > > > 
> > > > > If the delta is 1 tick then it is not supposed to stop it. Did you
> > > > > ever try to figure out WHY it is 1 tick?
> > > > > 
> > > > > There are two code pathes which can set it to basemono + TICK_NSEC:
> > > > > 
> > > > >         if (rcu_needs_cpu(basemono, &next_rcu) ||
> > > > >             arch_needs_cpu() || irq_work_needs_cpu()) {
> > > > >                 next_tick = basemono + TICK_NSEC;
> > > > >         } else {
> > > > >                 next_tmr = get_next_timer_interrupt(basejiff,
> > > > > basemono); ts->next_timer = next_tmr;
> > > > >                 /* Take the next rcu event into account */
> > > > >                 next_tick = next_rcu < next_tmr ? next_rcu : next_tmr;
> > > > >         }
> > > > > 
> > > > > Can you please figure out WHY the tick is requested to continue
> > > > > instead of blindly wreckaging the logic in that code?
> > > > 
> > > > Looks like the it hits in both cases during forced idle.
> > > > + Josh
> > > > + Paul
> > > > 
> > > > For the first case, it is always related to RCU. I found there are two
> > > > CONFIG options to avoid this undesired tick in idle loop.
> > > > 1. enable CONFIG_RCU_NOCB_CPU_ALL, offload to orcu kthreads
> > > > 2. or enable CONFIG_RCU_FAST_NO_HZ (enter dytick idle w/ rcu callback)
> > > > 
> > > > Either one works but my concern is that users may not realize the
> > > > intricate CONFIG_ options and how they translate into energy savings.
> > > > Consulted with Josh, it seems we could add a check here to recognize
> > > > the forced idle state and relax rcu_needs_cpu() to return false even it
> > > > has callbacks. Since we are blocking everybody for a short time (5 ticks
> > > > default). It should not impact synchronize and kfree rcu.
> > > 
> > > Right; as long as you're blocking *everybody*, and RCU priority boosting
> > > doesn't come into play (meaning a real-time task is waiting on RCU
> > > callbacks), then I don't see any harm in blocking RCU callbacks for a
> > > while.  You'd block completion of synchronize_rcu() and similar, as well
> > > as memory reclamation, but since you've blocked *every* CPU systemwide
> > > then that doesn't cause a problem.
> > 
> > True enough.  But how does RCU distinguish between this being a
> > normal idle cycle that might last indefinitely on the one hand and the
> > five-jiffy system-wide throttling on the other?  OK, maybe there is a
> > global variable that says that the just-now-starting idle period is
> > system-wide throttling.  But then what about the CPU that just went
> > idle 10 microseconds ago, and therefore left its timer tick running?
> > Fine and well, we could IPI it to wake it up and let it see that we
> > are now doing thermal throttling.  But then we presumably also have to
> > IPI it at the end of the thermal-throttling interval in order for it to
> > re-evaluate whether or not it should have the tick going.  :-/
> > 
> > On the one hand, I am sure that all of this can be made to work,
> > but simply having systems using thermal throttling enable either
> > CONFIG_RCU_NOCB_CPU_ALL or CONFIG_RCU_FAST_NO_HZ seems -way- simpler.
> > CONFIG_RCU_FAST_NO_HZ is probably the better choice for generic workloads,
> > but CONFIG_RCU_NOCB_CPU_ALL is the better choice for embedded workloads
> > where it is less likely that RCU callbacks will be posted with continuous
> > wild abandon.
> > 
> > Or am I missing something subtle here?
> 
> I agree that it seems preferable to make this require an existing RCU
> solution rather than adding more complexity to the RCU idle path.  One
> possible thing that may affect the choice of solution: this needs to
> idle *every* CPU, without leaving any CPU awake to handle callbacks or
> similar.

Fair point.  When in the five-jiffy throttling state, what can wake up
a CPU?  In an earlier version of this proposal, the answer was "nothing",
but maybe that has changed.

							Thanx, Paul