RE: [PATCH] [v4] x86, suspend: Save/restore extra MSR registers for suspend

["Doug Smythies" <dsmythies@telus.net>]

On 2015.11.01 08:50 Chen, Yu C wrote:
>> On 2015.10.10 19:27 Chen, Yu C wrote:
>>> On 2105.10.10 02:56 Doug Smythies wrote:
>>>
>>>>> The current version of the intel_pstate driver is incompatible with
>>>>> any use of Clock Modulation, always resulting in driving the target
>>>>> pstate to the minimum, regardless of load. The result is the
>>>>> apparent CPU frequency stuck at minimum * modulation percent.
>>>>
>>>>> The acpi-cpufreq driver works fine with Clock Modulation, resulting
>>>>> in desired frequency * modulation percent.
>>>>
>>
>>> [Yu] Why intel_pstate driver is incompatible with Clock Modulation?
>> 
>> It is simply how the current control algorithm responds to the scenario.
>> 
>> The problem is in intel_pstate_get_scaled_busy, here:
>> 
>>         /*
>>          * core_busy is the ratio of actual performance to max
>>          * max_pstate is the max non turbo pstate available
>>          * current_pstate was the pstate that was requested during
>>          *      the last sample period.
>>          *
>>          * We normalize core_busy, which was our actual percent
>>          * performance to what we requested during the last sample
>>          * period. The result will be a percentage of busy at a
>>          * specified pstate.
>>          */
>>         core_busy = cpu->sample.core_pct_busy;
>>         max_pstate = int_tofp(cpu->pstate.max_pstate);
>>         current_pstate = int_tofp(cpu->pstate.current_pstate);
>>         core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
>> 
>> With Clock Modulation enabled, the actual performance percent will always
>> be less than what was asked for, basically meaning current_pstate is much
>> less than what was asked for. Thus the algorithm will drive down the target
>> pstate regardless of load.
>> 
> [Yu] Do you mean, there is some problem with the normalization,and we should use 
> the actual pstate rather than the theoretical current_pstate, for example, 
> the pseudocode might looked like:
> 
> -  current_pstate = int_tofp(cpu->pstate.current_pstate);
> + current_pstate = int_tofp(cpu->pstate.current_pstat)*0.85;

I did not think of normalizing / compensating at this point.
That is a good idea.
Just for a test, I tried it and it seems to work well.
Before normalizing / compensating core_busy can be quite a small
for lesser clock modulation duty cycles, and so becomes a little
noisy afterwards. 

For my test, on an otherwise unaltered kernel v4.3 I did this:

diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
index aa33b92..97a90e1 100644
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -821,6 +821,7 @@ static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
        int32_t core_busy, max_pstate, current_pstate, sample_ratio;
        s64 duration_us;
        u32 sample_time;
+       u64 clock_modulation;

        /*
         * core_busy is the ratio of actual performance to max
@@ -836,6 +837,17 @@ static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
        core_busy = cpu->sample.core_pct_busy;
        max_pstate = int_tofp(cpu->pstate.max_pstate);
        current_pstate = int_tofp(cpu->pstate.current_pstate);
+
+//     rdmsrl(MSR_IA32_CLOCK_MODULATION, clock_modulation);
+       rdmsrl(MSR_IA32_THERM_CONTROL, clock_modulation);
+       if(clock_modulation && 0X10) {
+               clock_modulation = clock_modulation & 0x0F;
+               if(clock_modulation == 0) clock_modulation = 8;
+               core_busy = mul_fp(core_busy, int_tofp(0x10));
+               core_busy = div_fp(core_busy, int_tofp(clock_modulation));
+       }
+
        core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));

        /*