Re: [PATCH v2] cpufreq: schedutil: Serialize start against rate limit updates

[Rafael J. Wysocki (Intel)]

On Wed, Oct 7, 2026 at 5:59 PM Hui Su <sh_def@xxxxxxx> wrote:
>
> rate_limit_us_store() is called with the governor attribute set
> update_lock held, but sugov_start() updates freq_update_delay_ns from
> the same tunable without taking that lock.
>
> This allows the two paths to interleave as follows:
>
>         sugov_start()             rate_limit_us_store()
>
>         read old rate_limit_us
>                                   write new rate_limit_us
>                                   publish the new delay
>         publish the old delay
>
> The sysfs tunable then contains the new value while
> freq_update_delay_ns contains the old one. Unlike a transient torn
> read, the stale delay remains in effect until another rate limit update
> or governor restart.
>
> Serialize sugov_start() against sysfs stores by taking the existing
> governor attribute set lock around sugov_update_rate_limit_us(). Use
> gov_attr_set_lock() and gov_attr_set_unlock() to keep locking details
> within the attribute set interface. This keeps the fix off the scheduler
> hot path.
>
> Fixes: 9bdcb44e391d ("cpufreq: schedutil: New governor based on scheduler utilization data")
> Reviewed-by: Zhongqiu Han <zhongqiu.han@xxxxxxxxxxxxxxxx>
> Signed-off-by: Hui Su <sh_def@xxxxxxx>
> ---
> Changes in v2:
> - Wrap governor attribute set locking in helpers as suggested by
>   Zhongqiu Han.
>
> v1: https://lore.kernel.org/lkml/20260929103635.3780346-1-sh_def@xxxxxxx/
>
>  include/linux/cpufreq.h          | 10 ++++++++++
>  kernel/sched/cpufreq_schedutil.c |  4 ++++
>  2 files changed, 14 insertions(+)
>
> diff --git a/include/linux/cpufreq.h b/include/linux/cpufreq.h
> index 35ce665edfd8..5448eb0a0885 100644
> --- a/include/linux/cpufreq.h
> +++ b/include/linux/cpufreq.h
> @@ -694,6 +694,16 @@ static inline struct gov_attr_set *to_gov_attr_set(struct kobject *kobj)
>         return container_of(kobj, struct gov_attr_set, kobj);
>  }
>
> +static inline void gov_attr_set_lock(struct gov_attr_set *attr_set)
> +{
> +       mutex_lock(&attr_set->update_lock);
> +}
> +
> +static inline void gov_attr_set_unlock(struct gov_attr_set *attr_set)
> +{
> +       mutex_unlock(&attr_set->update_lock);
> +}

I'd rather not do this.

> +
>  void gov_attr_set_init(struct gov_attr_set *attr_set, struct list_head *list_node);
>  void gov_attr_set_get(struct gov_attr_set *attr_set, struct list_head *list_node);
>  unsigned int gov_attr_set_put(struct gov_attr_set *attr_set, struct list_head *list_node);
> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> index 49ccd6f1c185..b847311c307d 100644
> --- a/kernel/sched/cpufreq_schedutil.c
> +++ b/kernel/sched/cpufreq_schedutil.c
> @@ -861,7 +861,11 @@ static int sugov_start(struct cpufreq_policy *policy)
>         void (*uu)(struct update_util_data *data, u64 time, unsigned int flags);
>         unsigned int cpu;
>
> +       /* Serialize against rate_limit_us_store(). */
> +       gov_attr_set_lock(&sg_policy->tunables->attr_set);
>         sugov_update_rate_limit_us(sg_policy);
> +       gov_attr_set_unlock(&sg_policy->tunables->attr_set);

But instead, I'd add a wrapper around sugov_update_rate_limit_us() that would do

    guard(mutex)(&sg_policy->tunables->attr_set->update_lock);

before calling it and I'd invoke that wrapper from here.

Also, the comment should explain why the locking is needed which it
doesn't do now.

> +
>         sg_policy->last_freq_update_time        = 0;
>         sg_policy->next_freq                    = 0;
>         sg_policy->work_in_progress             = false;
>