Re: [PATCH v2] cpuidle: poll_state: Add time limit to poll_idle()

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Mar 12, 2018 at 10:36:27AM +0100, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> If poll_idle() is allowed to spin until need_resched() returns 'true',
> it may actually spin for a much longer time than expected by the idle
> governor, since set_tsk_need_resched() is not always called by the
> timer interrupt handler.  If that happens, the CPU may spend much
> more time than anticipated in the "polling" state.
> 
> To prevent that from happening, limit the time of the spinning loop
> in poll_idle().
> 
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
> 
> -> v2: After additional testing reduce POLL_IDLE_TIME_CHECK_COUNT to 1000.
> 
> ---
>  drivers/cpuidle/poll_state.c |   17 ++++++++++++++++-
>  1 file changed, 16 insertions(+), 1 deletion(-)
> 
> Index: linux-pm/drivers/cpuidle/poll_state.c
> ===================================================================
> --- linux-pm.orig/drivers/cpuidle/poll_state.c
> +++ linux-pm/drivers/cpuidle/poll_state.c
> @@ -5,16 +5,31 @@
>   */
>  
>  #include <linux/cpuidle.h>
> +#include <linux/ktime.h>
>  #include <linux/sched.h>
>  #include <linux/sched/idle.h>
>  
> +#define POLL_IDLE_TIME_CHECK_COUNT	1000
> +#define POLL_IDLE_TIME_LIMIT		(TICK_NSEC / 16)
> +
>  static int __cpuidle poll_idle(struct cpuidle_device *dev,
>  			       struct cpuidle_driver *drv, int index)
>  {
> +	ktime_t start = ktime_get();

I would recoomend not using ktime_get(), imagine the 'joy' if that
happens to be the HPET.

>  	local_irq_enable();
>  	if (!current_set_polling_and_test()) {
> +		unsigned int time_check_counter = 0;
> +
> +		while (!need_resched()) {
>  			cpu_relax();
> +			if (time_check_counter++ < POLL_IDLE_TIME_CHECK_COUNT)
> +				continue;
> +
> +			time_check_counter = 0;
> +			if (ktime_sub(ktime_get(), start) > POLL_IDLE_TIME_LIMIT)
> +				break;
> +		}
>  	}
>  	current_clr_polling();

Since the idle loop is strictly per-cpu, you can use regular
sched_clock() here. Something like:

	u64 start = sched_clock();

	local_irq_enable();
	if (!current_set_polling_and_test()) {
		while (!need_resched()) {
			cpu_relax();

			if (sched_clock() - start > POLL_IDLE_TIME_LIMIT)
				break;
		}
	}
	current_clr_polling();

On x86 we don't have to use that time_check_counter thing, sched_clock()
is really cheap, not sure if it makes sense on other platforms.