Re: [PATCH] pmdomain: governor: Consider CPU latency tolerance from pm_domain_cpu_gov

[Ulf Hansson <ulf.hansson@linaro.org>]

On Tue, 8 Jul 2025 at 07:15, Maulik Shah <maulik.shah@oss.qualcomm.com> wrote:
>
> pm_domain_cpu_gov is selecting a cluster idle state but does not consider
> latency tolerance of child CPUs. This results in deeper cluster idle state
> whose latency does not meet latency tolerance requirement.
>
> Select deeper idle state only if global and device latency tolerance of all
> child CPUs meet.
>
> Test results on SM8750 with 300 usec PM-QoS on CPU0 which is less than
> domain idle state entry (2150) + exit (1983) usec latency mentioned in
> devicetree, demonstrate the issue.
>
>         # echo 300 > /sys/devices/system/cpu/cpu0/power/pm_qos_resume_latency_us
>
> Before: (Usage is incrementing)
> ======
>         # cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states
>         State          Time Spent(ms) Usage      Rejected   Above      Below
>         S0             29817          537        8          270        0
>
>         # cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states
>         State          Time Spent(ms) Usage      Rejected   Above      Below
>         S0             30348          542        8          271        0
>
> After: (Usage is not incrementing due to latency tolerance)
> ======
>         # cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states
>         State          Time Spent(ms) Usage      Rejected   Above      Below
>         S0             39319          626        14         307        0
>
>         # cat /sys/kernel/debug/pm_genpd/power-domain-cluster0/idle_states
>         State          Time Spent(ms) Usage      Rejected   Above      Below
>         S0             39319          626        14         307        0
>
> Signed-off-by: Maulik Shah <maulik.shah@oss.qualcomm.com>

Thanks for posting this! I am actually carrying a local patch in my
tree for this, just haven't reached the point of posting it. :-)

> ---
>  drivers/pmdomain/governor.c | 18 ++++++++++++++++--
>  1 file changed, 16 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/pmdomain/governor.c b/drivers/pmdomain/governor.c
> index c1e148657c873a6b5b4d9c0f058d54cb020c56e2..d5ac4c1b5b5a432f0072209d17379e58ec891202 100644
> --- a/drivers/pmdomain/governor.c
> +++ b/drivers/pmdomain/governor.c
> @@ -8,6 +8,7 @@
>  #include <linux/pm_domain.h>
>  #include <linux/pm_qos.h>
>  #include <linux/hrtimer.h>
> +#include <linux/cpu.h>
>  #include <linux/cpuidle.h>
>  #include <linux/cpumask.h>
>  #include <linux/ktime.h>
> @@ -349,6 +350,8 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd)
>         struct cpuidle_device *dev;
>         ktime_t domain_wakeup, next_hrtimer;
>         ktime_t now = ktime_get();
> +       struct device *device;

Nitpick: Maybe rename to "cpu_dev" to better distinguish between the
cpuidle_device *dev, above?

> +       s64 cpu_constraint, global_constraint;
>         s64 idle_duration_ns;
>         int cpu, i;
>
> @@ -359,6 +362,7 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd)
>         if (!(genpd->flags & GENPD_FLAG_CPU_DOMAIN))
>                 return true;
>
> +       global_constraint = cpu_latency_qos_limit();
>         /*
>          * Find the next wakeup for any of the online CPUs within the PM domain
>          * and its subdomains. Note, we only need the genpd->cpus, as it already
> @@ -372,8 +376,16 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd)
>                         if (ktime_before(next_hrtimer, domain_wakeup))
>                                 domain_wakeup = next_hrtimer;
>                 }
> +
> +               device = get_cpu_device(cpu);
> +               if (device) {
> +                       cpu_constraint = dev_pm_qos_read_value(device, DEV_PM_QOS_RESUME_LATENCY);

We should be able to use dev_pm_qos_raw_resume_latency() here, similar
to how cpuidle_governor_latency_req() does it. I think it's better as
it avoids acquiring/releasing the spinlock.

> +                       if (cpu_constraint < global_constraint)
> +                               global_constraint = cpu_constraint;
> +               }
>         }
>
> +       global_constraint *= NSEC_PER_USEC;
>         /* The minimum idle duration is from now - until the next wakeup. */
>         idle_duration_ns = ktime_to_ns(ktime_sub(domain_wakeup, now));
>         if (idle_duration_ns <= 0)
> @@ -389,8 +401,10 @@ static bool cpu_power_down_ok(struct dev_pm_domain *pd)
>          */
>         i = genpd->state_idx;
>         do {
> -               if (idle_duration_ns >= (genpd->states[i].residency_ns +
> -                   genpd->states[i].power_off_latency_ns)) {
> +               if ((idle_duration_ns >= (genpd->states[i].residency_ns +
> +                   genpd->states[i].power_off_latency_ns)) &&
> +                   (global_constraint >= (genpd->states[i].power_on_latency_ns +
> +                   genpd->states[i].power_off_latency_ns))) {
>                         genpd->state_idx = i;
>                         genpd->gd->last_enter = now;
>                         genpd->gd->reflect_residency = true;
>
> ---
> base-commit: 50c8770a42faf8b1c7abe93e7c114337f580a97d
> change-id: 20250703-pmdomain_qos-25b8dbb623ea
>
> Best regards,
> --
> Maulik Shah <maulik.shah@oss.qualcomm.com>
>

Besides the minor things above, this looks good to me.

Kind regards
Uffe