Re: [PATCH 2/2] cpufreq: Update CPU capacity reduction in store_scaling_max_freq()

[Lukasz Luba <lukasz.luba@arm.com>]

+CC Daniel

On 10/10/22 11:22, Vincent Guittot wrote:
> On Mon, 10 Oct 2022 at 12:12, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>>
>>
>> On 10/10/22 10:32, Vincent Guittot wrote:
>>> On Mon, 10 Oct 2022 at 11:30, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 10/10/22 10:15, Vincent Guittot wrote:
>>>>> On Mon, 10 Oct 2022 at 11:02, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 10/10/22 06:39, Viresh Kumar wrote:
>>>>>>> Would be good to always CC Scheduler maintainers for such a patch.
>>>>>>
>>>>>> Agree, I'll do that.
>>>>>>
>>>>>>>
>>>>>>> On 30-09-22, 10:48, Lukasz Luba wrote:
>>>>>>>> When the new max frequency value is stored, the task scheduler must
>>>>>>>> know about it. The scheduler uses the CPUs capacity information in the
>>>>>>>> task placement. Use the existing mechanism which provides information
>>>>>>>> about reduced CPU capacity to the scheduler due to thermal capping.
>>>>>>>>
>>>>>>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
>>>>>>>> ---
>>>>>>>>      drivers/cpufreq/cpufreq.c | 18 +++++++++++++++++-
>>>>>>>>      1 file changed, 17 insertions(+), 1 deletion(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
>>>>>>>> index 1f8b93f42c76..205d9ea9c023 100644
>>>>>>>> --- a/drivers/cpufreq/cpufreq.c
>>>>>>>> +++ b/drivers/cpufreq/cpufreq.c
>>>>>>>> @@ -27,6 +27,7 @@
>>>>>>>>      #include <linux/slab.h>
>>>>>>>>      #include <linux/suspend.h>
>>>>>>>>      #include <linux/syscore_ops.h>
>>>>>>>> +#include <linux/thermal.h>
>>>>>>>>      #include <linux/tick.h>
>>>>>>>>      #include <linux/units.h>
>>>>>>>>      #include <trace/events/power.h>
>>>>>>>> @@ -718,6 +719,8 @@ static ssize_t show_scaling_cur_freq(struct cpufreq_policy *policy, char *buf)
>>>>>>>>      static ssize_t store_scaling_max_freq
>>>>>>>>      (struct cpufreq_policy *policy, const char *buf, size_t count)
>>>>>>>>      {
>>>>>>>> +    unsigned int frequency;
>>>>>>>> +    struct cpumask *cpus;
>>>>>>>>         unsigned long val;
>>>>>>>>         int ret;
>>>>>>>>
>>>>>>>> @@ -726,7 +729,20 @@ static ssize_t store_scaling_max_freq
>>>>>>>>                 return -EINVAL;
>>>>>>>>
>>>>>>>>         ret = freq_qos_update_request(policy->max_freq_req, val);
>>>>>>>> -    return ret >= 0 ? count : ret;
>>>>>>>> +    if (ret >= 0) {
>>>>>>>> +            /*
>>>>>>>> +             * Make sure that the task scheduler sees these CPUs
>>>>>>>> +             * capacity reduction. Use the thermal pressure mechanism
>>>>>>>> +             * to propagate this information to the scheduler.
>>>>>>>> +             */
>>>>>>>> +            cpus = policy->related_cpus;
>>>>>>>
>>>>>>> No need of this, just use related_cpus directly.
>>>>>>>
>>>>>>>> +            frequency = __resolve_freq(policy, val, CPUFREQ_RELATION_HE);
>>>>>>>> +            arch_update_thermal_pressure(cpus, frequency);
>>>>>>>
>>>>>>> I wonder if using the thermal-pressure API here is the right thing to
>>>>>>> do. It is a change coming from User, which may or may not be
>>>>>>> thermal-related.
>>>>>>
>>>>>> Yes, I thought the same. Thermal-pressure name might be not the
>>>>>> best for covering this use case. I have been thinking about this
>>>>>> thermal pressure mechanism for a while, since there are other
>>>>>> use cases like PowerCap DTPM which also reduces CPU capacity
>>>>>> because of power policy from user-space. We don't notify
>>>>>> the scheduler about it. There might be also an issue with virtual
>>>>>> guest OS and how that kernel 'sees' the capacity of CPUs.
>>>>>> We might try to use this 'thermal-pressure' in the guest kernel
>>>>>> to notify about available CPU capacity (just a proposal, not
>>>>>> even an RFC, since we are missing requirements, but issues where
>>>>>> discussed on LPC 2022 on ChromeOS+Android_guest)
>>>>>
>>>>> The User space setting scaling_max_freq is a long scale event and it
>>>>> should be considered as a new running environnement instead of a
>>>>> transient event. I would suggest updating the EM is and capacity orig
>>>>> of the system in this case. Similarly, we rebuild sched_domain with a
>>>>> cpu hotplug. scaling_max_freq interface should not be used to do any
>>>>> kind of dynamic scaling.
>>>>
>>>> I tend to agree, but the EM capacity would be only used in part of EAS
>>>> code. The whole fair.c view to the capacity_of() (RT + DL + irq +
>>>> thermal_pressure) would be still wrong in other parts, e.g.
>>>> select_idle_sibling() and load balance.
>>>>
>>>> When we get this powerhint we might be already in overutilied state
>>>> so EAS is disabled. IMO other mechanisms in the task scheduler
>>>> should be also aware of that capacity reduction.
>>>
>>> That's why I also mentioned the capacity_orig
>>
>> Well, I think this is a bit more complex. Thermal framework governor
>> reduces the perf IDs from top in the freq asc table and keeps that
>> in the statistics in sysfs. It also updates the thermal pressure signal.
>> When we rebuild the capacity of CPUs and make the capacity_orig smaller,
>> the capacity_of would still have the thermal framework reduced capacity
>> in there. We would end up with too small CPU capacity due to this
>> subtraction in capacity_of.
> 
> That's why using user space interface should not be used to do dynamic scaling.
> I still think that user space interface is not the right interface
> 
>>
>> Ideally, I would see a mechanism which is aware of this performance
>> reduction reason:
>> 1. thermal capping
>> 2. power capping (from DTPM)
>> 3. max freq reduction by user space
> 
> Yes for thermal and power capping  but no for user space
> 
>>
>> That common place would figure and maintain the context for the
>> requested capacity reduction.
>>
>> BTW, those Android user space max freq requests are not that long,
>> mostly due to camera capturing (you can see a few in this file,
>> e.g. [1]).
> 
> Why are they doing this ?
> This doesn't seem to be the correct interface to use. It seems to do
> some power budget and they should use the right interface for this

Yes, I agree. I have sent explanation with this to Peter's emails.
Daniel tries to give them a better interface: DTPM, but also would
suffer the same issue of capacity reduction for this short time.

We have a few discussions about it, also Daniel presented on a few
LPC those issues.