Re: [PATCH V3] cpuidle: Handle tick_broadcast_enter() failure gracefully

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

On Friday, May 08, 2015 04:18:02 PM Rafael J. Wysocki wrote:
> On Friday, May 08, 2015 01:05:32 PM Preeti U Murthy wrote:
> > When a CPU has to enter an idle state where tick stops, it makes a call
> > to tick_broadcast_enter(). The call will fail if this CPU is the
> > broadcast CPU. Today, under such a circumstance, the arch cpuidle code
> > handles this CPU.  This is not convincing because not only do we not
> > know what the arch cpuidle code does, but we also do not account for the
> > idle state residency time and usage of such a CPU.
> > 
> > This scenario can be handled better by simply choosing an idle state
> > where in ticks do not stop. To accommodate this change move the setting
> > of runqueue idle state from the core to the cpuidle driver, else the
> > rq->idle_state will be set wrong.
> > 
> > Signed-off-by: Preeti U Murthy <preeti@linux.vnet.ibm.com>
> > ---
> > Changes from V2: https://lkml.org/lkml/2015/5/7/78
> > Introduce a function in cpuidle core to select an idle state where ticks do not
> > stop rather than going through the governors.
> > 
> > Changes from V1: https://lkml.org/lkml/2015/5/7/24
> > Rebased on the latest linux-pm/bleeding-edge branch
> > 
> >  drivers/cpuidle/cpuidle.c |   45 +++++++++++++++++++++++++++++++++++++++++++--
> >  include/linux/sched.h     |   16 ++++++++++++++++
> >  kernel/sched/core.c       |   17 +++++++++++++++++
> >  kernel/sched/fair.c       |    2 +-
> >  kernel/sched/idle.c       |    6 ------
> >  kernel/sched/sched.h      |   24 ------------------------
> >  6 files changed, 77 insertions(+), 33 deletions(-)
> > 
> > diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c
> > index 8c24f95..d1af760 100644
> > --- a/drivers/cpuidle/cpuidle.c
> > +++ b/drivers/cpuidle/cpuidle.c
> > @@ -21,6 +21,7 @@
> >  #include <linux/module.h>
> >  #include <linux/suspend.h>
> >  #include <linux/tick.h>
> > +#include <linux/sched.h>
> >  #include <trace/events/power.h>
> >  
> >  #include "cpuidle.h"
> > @@ -146,6 +147,36 @@ int cpuidle_enter_freeze(struct cpuidle_driver *drv, struct cpuidle_device *dev)
> >  	return index;
> >  }
> >  
> > +/*
> > + * find_tick_valid_state - select a state where tick does not stop
> > + * @dev: cpuidle device for this cpu
> > + * @drv: cpuidle driver for this cpu
> > + */
> > +static int find_tick_valid_state(struct cpuidle_device *dev,
> > +				struct cpuidle_driver *drv)
> > +{
> > +	int i, ret = -1;
> > +
> > +	for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
> > +		struct cpuidle_state *s = &drv->states[i];
> > +		struct cpuidle_state_usage *su = &dev->states_usage[i];
> > +
> > +		/*
> > +		 * We do not explicitly check for latency requirement
> > +		 * since it is safe to assume that only shallower idle
> > +		 * states will have the CPUIDLE_FLAG_TIMER_STOP bit
> > +		 * cleared and they will invariably meet the latency
> > +		 * requirement.
> > +		 */
> > +		if (s->disabled || su->disable ||
> > +			(s->flags & CPUIDLE_FLAG_TIMER_STOP))
> > +			continue;
> > +
> > +		ret = i;
> > +	}
> > +	return ret;
> > +}
> > +
> >  /**
> >   * cpuidle_enter_state - enter the state and update stats
> >   * @dev: cpuidle device for this cpu
> > @@ -168,10 +199,17 @@ int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
> >  	 * CPU as a broadcast timer, this call may fail if it is not available.
> >  	 */
> >  	if (broadcast && tick_broadcast_enter()) {
> > -		default_idle_call();
> > -		return -EBUSY;
> > +		index = find_tick_valid_state(dev, drv);
> 
> Well, the new state needs to be deeper

I should have said "shallower", sorry about that.

The state chosen by the governor satisfies certain latency requirements and we
can't violate those by choosing a deeper state here.

But the patch I sent actually did the right thing. :-)

> than the old one or you may violate the governor's choice and this doesn't
> guarantee that.
> 
> Also I don't quite see a reason to duplicate the find_deepest_state() functionality
> here.
> 
> > +		if (index < 0) {
> > +			default_idle_call();
> > +			return -EBUSY;
> > +		}
> > +		target_state = &drv->states[index];
> >  	}
> >  
> > +	/* Take note of the planned idle state. */
> > +	idle_set_state(smp_processor_id(), target_state);
> 
> And I wouldn't do this either.
> 
> The behavior here is pretty much as though the driver demoted the state chosen
> by the governor and we don't call idle_set_state() again in those cases.
> 
> > +
> >  	trace_cpu_idle_rcuidle(index, dev->cpu);
> >  	time_start = ktime_get();
> 
> Overall, something like the patch below (untested) should work I suppose?
> 
> ---
>  drivers/cpuidle/cpuidle.c |   21 ++++++++++++++-------
>  1 file changed, 14 insertions(+), 7 deletions(-)
> 
> Index: linux-pm/drivers/cpuidle/cpuidle.c
> ===================================================================
> --- linux-pm.orig/drivers/cpuidle/cpuidle.c
> +++ linux-pm/drivers/cpuidle/cpuidle.c
> @@ -73,17 +73,19 @@ int cpuidle_play_dead(void)
>  }
>  
>  static int find_deepest_state(struct cpuidle_driver *drv,
> -			      struct cpuidle_device *dev, bool freeze)
> +			      struct cpuidle_device *dev, bool freeze,
> +			      int limit, unsigned int flags_to_avoid)
>  {
>  	unsigned int latency_req = 0;
>  	int i, ret = freeze ? -1 : CPUIDLE_DRIVER_STATE_START - 1;
>  
> -	for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
> +	for (i = CPUIDLE_DRIVER_STATE_START; i < limit; i++) {
>  		struct cpuidle_state *s = &drv->states[i];
>  		struct cpuidle_state_usage *su = &dev->states_usage[i];
>  
>  		if (s->disabled || su->disable || s->exit_latency <= latency_req
> -		    || (freeze && !s->enter_freeze))
> +		    || (freeze && !s->enter_freeze)
> +		    || (s->flags & flags_to_avoid))
>  			continue;
>  
>  		latency_req = s->exit_latency;
> @@ -100,7 +102,7 @@ static int find_deepest_state(struct cpu
>  int cpuidle_find_deepest_state(struct cpuidle_driver *drv,
>  			       struct cpuidle_device *dev)
>  {
> -	return find_deepest_state(drv, dev, false);
> +	return find_deepest_state(drv, dev, false, drv->state_count, 0);
>  }
>  
>  static void enter_freeze_proper(struct cpuidle_driver *drv,
> @@ -139,7 +141,7 @@ int cpuidle_enter_freeze(struct cpuidle_
>  	 * that interrupts won't be enabled when it exits and allows the tick to
>  	 * be frozen safely.
>  	 */
> -	index = find_deepest_state(drv, dev, true);
> +	index = find_deepest_state(drv, dev, true, drv->state_count, 0);
>  	if (index >= 0)
>  		enter_freeze_proper(drv, dev, index);
>  
> @@ -168,8 +170,13 @@ int cpuidle_enter_state(struct cpuidle_d
>  	 * CPU as a broadcast timer, this call may fail if it is not available.
>  	 */
>  	if (broadcast && tick_broadcast_enter()) {
> -		default_idle_call();
> -		return -EBUSY;
> +		index = find_deepest_state(drv, dev, false, index,
> +					   CPUIDLE_FLAG_TIMER_STOP);
> +		if (index < 0) {
> +			default_idle_call();
> +			return -EBUSY;
> +		}
> +		target_state = &drv->states[index];
>  	}
>  
>  	trace_cpu_idle_rcuidle(index, dev->cpu);
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.