Re: [PATCH 2/2] sched/cpufreq: Update schedutil's DVFS request to reach the boost frequencies

On 8/6/2026 8:43 PM, Vincent Guittot wrote:
> On Thu, 6 Aug 2026 at 11:02, Christian Loehle <[email protected]> wrote:
>> On 8/6/26 05:42, Sibi Sankar wrote:
>>> capacity_freq_ref, exposed to schedutil via get_capacity_ref_freq(),
>>> was introduced by commit 9942cb22ea45 ("sched/topology: Add a new
>>> arch_scale_freq_ref() method") as a fixed anchor that does not move at
>>> runtime. However, schedutil uses that same fixed anchor as the reference
>>> plugged into map_util_freq() which saturates exactly at capacity_freq_ref.
>>>
>>> As a result, a system with cpufreq boost enabled effectively never runs at
>>> boost frequencies under schedutil-governed load. Fix this by plugging in
>>> policy-max into the map_util_freq equation, so that the DVFS requests
>>> translates to the actual cpufreq driver ceiling.
>>>
>>> Signed-off-by: Sibi Sankar <[email protected]>
>>> ---
>>>   kernel/sched/cpufreq_schedutil.c | 13 ++++++++++---
>>>   1 file changed, 10 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
>>> index a1782755efcc..dcefbeaa0702 100644
>>> --- a/kernel/sched/cpufreq_schedutil.c
>>> +++ b/kernel/sched/cpufreq_schedutil.c
>>> @@ -195,10 +195,17 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
>>>                                  unsigned long util, unsigned long max)
>>>   {
>>>        struct cpufreq_policy *policy = sg_policy->policy;
>>> -     unsigned int freq;
>>> +     unsigned int freq, ref;
>>>
>>> -     freq = get_capacity_ref_freq(policy);
>>> -     freq = map_util_freq(util, freq, max);
>>> +     ref = get_capacity_ref_freq(policy);
>>> +
>>> +     /*
>>> +      * That fixed anchor governs how utilization is interpreted, but
>>> +      * the DVFS request is free to target the current policy ceiling.
>>> +      * Using ref alone would saturate the util->freq map at ref so
>>> +      * use policy->max to reach boost frequencies.
>> I'm not sure those two statements are compatible with the implementation below?
>> util / max is expressed in the capacity scale established using ref, multiplying
>> that ratio by policy->max changes the interpretation of every util value.
>>
>>
>>> +      */
>>> +     freq = map_util_freq(util, max(ref, READ_ONCE(policy->max)), max);
>> Isn't the underlying problem that schedutil can't handle requests above capacity 1024?
>>
>> effective_cpu_util() caps util at max, and sugov_effective_cpu_perf() applies
>> the 25% headroom before clipping the result back to that same value.
>> Therefore the input here cannot exceed 1024, and mapping it against ref
>> can never request a frequency above ref.
>> Using policy->max makes boost reachable, but also stretches the complete frequency range.
>> For example, with ref = 4454400, policy->max = 4723200, and effective util 640, the request
>> changes from 2784000 to 2952000.
>>
>> Making 'boost frequencies' truly compatible with schedutil (or PELT/CAS for that matter)
>> is a discussion that is yet to be had?
>> See also
>> https://lore.kernel.org/lkml/[email protected]/
> cpufreq_pressure and freq_qos_update_request could handle the case
> where boost freq is used as the ref freq when setting capacity but
> disabled. I can't remember if everything was already in place when we
> discussed this last time

Christian/Vincent,
Thanks, will look at how acpi-cpufreq/ccp-cpufreq in the x86 world handle
this and try out some of the suggestions in the link and get back.

>