[PATCH v1] cpuidle: Add sanity check for exit latency and target residency

[PATCH v1] cpuidle: Add sanity check for exit latency and target residency

[Rafael J. Wysocki]

From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Make __cpuidle_driver_init() sanitize CPU idle states so that the exit
latency of a given state is not greater than its target residency which
would break cpuidle assumptions.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---
 drivers/cpuidle/driver.c |   11 +++++++++++
 1 file changed, 11 insertions(+)

--- a/drivers/cpuidle/driver.c
+++ b/drivers/cpuidle/driver.c
@@ -193,6 +193,16 @@ static void __cpuidle_driver_init(struct
 			s->exit_latency_ns =  0;
 		else
 			s->exit_latency = div_u64(s->exit_latency_ns, NSEC_PER_USEC);
+
+		/*
+		 * Ensure that the exit latency of a CPU idle state does not
+		 * exceed its target residency which is assumed in cpuidle in
+		 * multiple places.
+		 */
+		if (s->exit_latency_ns > s->target_residency_ns) {
+			s->target_residency_ns = s->exit_latency_ns;
+			s->target_residency = s->exit_latency;
+		}
 	}
 }
 

Re: [PATCH v1] cpuidle: Add sanity check for exit latency and target residency

[Artem Bityutskiy]

On Fri, 2025-11-07 at 19:19 +0100, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> Make __cpuidle_driver_init() sanitize CPU idle states so that the exit
> latency of a given state is not greater than its target residency which
> would break cpuidle assumptions.
> 
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
>  drivers/cpuidle/driver.c |   11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> --- a/drivers/cpuidle/driver.c
> +++ b/drivers/cpuidle/driver.c
> @@ -193,6 +193,16 @@ static void __cpuidle_driver_init(struct
>  			s->exit_latency_ns =  0;
>  		else
>  			s->exit_latency = div_u64(s->exit_latency_ns, NSEC_PER_USEC);
> +
> +		/*
> +		 * Ensure that the exit latency of a CPU idle state does not
> +		 * exceed its target residency which is assumed in cpuidle in
> +		 * multiple places.
> +		 */
> +		if (s->exit_latency_ns > s->target_residency_ns) {
> +			s->target_residency_ns = s->exit_latency_ns;
> +			s->target_residency = s->exit_latency;
> +		}

I suggest to error out instead of capping it. Because as soon as you
cap it, you may end up with the target residency of the next C-state. 


Just erroring out is very explicit, no surprises, and the table
provider will have to just fix the table.


Thanks, Artem.

Re: [PATCH v1] cpuidle: Add sanity check for exit latency and target residency

[Rafael J. Wysocki]

On Fri, Nov 7, 2025 at 7:47 PM Artem Bityutskiy
<artem.bityutskiy@linux.intel.com> wrote:
>
> On Fri, 2025-11-07 at 19:19 +0100, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> >
> > Make __cpuidle_driver_init() sanitize CPU idle states so that the exit
> > latency of a given state is not greater than its target residency which
> > would break cpuidle assumptions.
> >
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > ---
> >  drivers/cpuidle/driver.c |   11 +++++++++++
> >  1 file changed, 11 insertions(+)
> >
> > --- a/drivers/cpuidle/driver.c
> > +++ b/drivers/cpuidle/driver.c
> > @@ -193,6 +193,16 @@ static void __cpuidle_driver_init(struct
> >                       s->exit_latency_ns =  0;
> >               else
> >                       s->exit_latency = div_u64(s->exit_latency_ns, NSEC_PER_USEC);
> > +
> > +             /*
> > +              * Ensure that the exit latency of a CPU idle state does not
> > +              * exceed its target residency which is assumed in cpuidle in
> > +              * multiple places.
> > +              */
> > +             if (s->exit_latency_ns > s->target_residency_ns) {
> > +                     s->target_residency_ns = s->exit_latency_ns;
> > +                     s->target_residency = s->exit_latency;
> > +             }
>
> I suggest to error out instead of capping it. Because as soon as you
> cap it, you may end up with the target residency of the next C-state.

Good point.

> Just erroring out is very explicit, no surprises, and the table
> provider will have to just fix the table.

Alternatively, the state may just be dropped, but let's try with
failing to begin with.