Re: [PATCH v1 1/4] cpuidle: teo: Add polling flag check to early return path

[Christian Loehle <christian.loehle@arm.com>]

On 1/10/25 12:53, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> After commit 6da8f9ba5a87 ("cpuidle: teo: Skip tick_nohz_get_sleep_length()
> call in some cases") the teo governor behaves a bit differently on
> systems where idle state 0 is a "polling" state (that is, it is not
> really an idle state, but a loop continuously executed by the CPU).
> Namely, on such systems it skips the tick_nohz_get_sleep_length() call
> if the target residency of the current candidate idle state is small
> enough.
> 
> However, if state 0 itself was to be returned, it would be returned
> right away without calling tick_nohz_get_sleep_length() even on systems
> where it was not a "polling" state until commit 4b20b07ce72f ("cpuidle:
> teo: Don't count non-existent intercepts") that attempted to fix this
> problem.
> 
> Unfortunately, commit 4b20b07ce72f has made the governor always call
> tick_nohz_get_sleep_length() when about to return state 0 early, even
> if that state is a "polling" one, which is inconsistent and defeats
> the purpose of commit 6da8f9ba5a87 in that case.
> 
> Address this by adding a CPUIDLE_FLAG_POLLING check to the path where
> state 0 is returned early to prevent tick_nohz_get_sleep_length() from
> being called if it is a "polling" state.
> 
> Fixes: 4b20b07ce72f ("cpuidle: teo: Don't count non-existent intercepts")
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
>  drivers/cpuidle/governors/teo.c |    3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> --- a/drivers/cpuidle/governors/teo.c
> +++ b/drivers/cpuidle/governors/teo.c
> @@ -422,7 +422,8 @@
>  			first_suitable_idx = i;
>  		}
>  	}
> -	if (!idx && prev_intercept_idx) {
> +	if (!idx && prev_intercept_idx &&
> +	    !(drv->states[0].flags & CPUIDLE_FLAG_POLLING)) {
>  		/*
>  		 * We have to query the sleep length here otherwise we don't
>  		 * know after wakeup if our guess was correct.
> 
> 
> 

But then you do run into the issue of intercepts not being detected if
state0 is the right choice, don't you?
This would then enable intercept-detection only for <50% of the time,
another option is to not allow intercepts selecting a polling state, but
there were recent complaints about this exact behavior from Aboorva (+TO).
They don't have a low-latency non-polling state.

https://lore.kernel.org/lkml/20240809073120.250974-1-aboorvad@linux.ibm.com/

Regards,
Christian