lkml.org 
[lkml]   [2015]   [Nov]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
SubjectRE: [PATCH] [v4] x86, suspend: Save/restore extra MSR registers for suspend
Date

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));

/*





\
 
 \ /
  Last update: 2015-11-06 16:41    [W:0.082 / U:0.020 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site